The Automatic palette counts flat art's colours by area (a colour needs
about 1% of the picture, 0.5% at Colors 100) and fits that count with
median cut and area-weighted k-means. A stem, a highlight, or a near
white beside a pale background (within 24 levels, so in one box) was
averaged into a large neighbour; mixed-colour merging and Noise could
then remove what was left.
On the flat route only, after the count is fitted, rare_colours finds
the colours too rare or too close for it and keeps those that are
distinct (at least 2.5 in Oklab x 100 from every palette colour and from
the main colour each stands for), flat (mostly one histogram bin) and
drawn (a 4-connected region of their own unblended pixels of at least 8
pixels, or three of at least 4); a colour lying between two others, as
anti-aliasing mixes them, must also have an inside. At most six, largest
region first. They stay fixed while the other colours are refitted
around them; merge_mixed leaves them alone; denoise_keeping keeps their
regions of 4 pixels or more whatever Noise is (smaller specks still go).
Full Tone, Limited and Automatic's continuous-tone route are unchanged,
and denoise without kept labels is as before.
Methods cited in the code: median cut (Heckbert 1982), k-means (Lloyd
1982), Oklab (Ottosson 2020), salient-colour preservation (Fukuda et al.
2019, doi:10.1117/12.2521607). The rule and its numbers are VectorCraft's
own; Adobe does not document how its palettes are chosen.
Tests (code-drawn): a rare stem or patch over 2-8 flat colours, hard
and anti-aliased at 1x and 2x; near-white highlights on #FFF4E0 kept and
still paths at Noise 16; a 3 px #909090 stem on #808080; scattered
single pixels of the same areas add no colour; denoise_keeping at Noise
0-32.
Measured on 66 synthetic flat logos (22 designs x hard 1x, anti-aliased
1x and 2x; local inputs): every key colour and no spurious one in 66
(was 39); missing key colours 0 (was 43), spurious 0 (was 2), merged
pairs 0 (was 46); rendered dE00 0.348 -> 0.138. Four public-domain scan
crops take the continuous-tone route and are unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018tBcdwNhptqtmX32JCGFzB
The public documentation says Image Trace warns about images larger
than 1500 pixels on a side but traces them, only more slowly. VectorCraft
traced them silently, and refused any trace of more than 2,000,000
anchors, which a 4000 x 3000 photo at High Fidelity Photo nearly reaches
(1.9 million on a code-drawn photo) and a noisier one passes.
- imageTrace.make and makeAndExpand return `warnings`: one for an image
over 1500 px on a side (traced, but more slowly; a lower-fidelity
preset or a smaller copy is quicker), one for an image over 64 MP
that was traced from a scaled-down copy. The app shows the first in
the status bar.
- MAX_TRACE_ANCHORS is 8,000,000 natively (2,000,000 on wasm, where a
page has at most 4 GB). Measured through imageTrace.make on
code-drawn photos: about 90 bytes an anchor kept in the document;
7.9 million anchors kept 685 MB with an 876 MB peak (4000 x 3000,
Full Tone at Colors 100, Paths 100, Noise 1, 37 s).
- The refusal keeps its translated advice: raise Noise, lower Paths or
Colors, or pick a lower-fidelity preset.
Tests: a_large_image_traces_with_a_warning (1600 x 40 and 1500 x 8
code-drawn images, make and make-and-expand),
trace_warnings_name_a_large_or_scaled_down_image, a compile-time check
of the limit's range, and a UI test that the status bar shows the note.
Source:
https://helpx.adobe.com/illustrator/using/image-trace-results-optimization.html
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018tBcdwNhptqtmX32JCGFzB
Raster::new checked its RGBA length with assert_eq!, so bytes of the
wrong length (from a file, a plug-in or a control message) would abort
instead of reporting an error. It now returns Result<Raster, TraceError>
and gives TraceError::Decode naming the byte count and the size (the
existing variant, whose text is already translated, instead of a new
message every language would need). The size product is checked for
overflow.
Raster::pixel indexed the bytes directly; it now gives transparent
black outside the raster, or past the end of a raster whose public
fields were set to a size its bytes don't have (which mosaic reads).
The public signature changes: every caller in the workspace was a test
(trace unit tests, Flat Logo tests and the quality harness), updated to
unwrap. No shipped code built a Raster with new; decode, decode_to_fit
and from_fn always make the right length, and trace_within already
refuses a raster whose fields disagree.
Test (code-drawn): a_raster_of_the_wrong_length_is_an_error_not_a_crash
(short, long, empty, zero-width and overflowing sizes; pixel, trace and
mosaic on a raster whose fields disagree).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018tBcdwNhptqtmX32JCGFzB
Place junction clusters at their deepest pixel and prune terminal branches
by significance beyond their maximal discs. Merge overlapping junctions
and split crossings with continuing outer arms, then join degree-two runs.
Bound nodes, branches and simplification work, using checked data access.
Keep whole-blob selection, ink refinement, cap detection and fitting rules.
Test junction degrees, crossings, 24-angle lines, pruning, graph caps and
random-blob determinism. Record improved junction degrees and remaining
spur and extra-junction limitations in the centre-lines measurements.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: OpenAI <noreply@openai.com>
Compute exact squared Euclidean distances per blob with separable parabola
lower envelopes. Thin with bounded Guo–Hall active sets and retain diagonal
ends during staircase cleanup. Keep the whole-blob width rule and existing
graph, ink, cap and fitting rules, with checked access in the moved module.
Verify EDT against brute force, active sets against plain rounds, and both
diagonal slopes for length and connectivity. Record the centre-lines rerun
and its remaining limitations without changing feature scores.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: OpenAI <noreply@openai.com>
Measure Create Strokes on exact code-drawn scenes in aa and hard variants,
including centre/width error, topology, ends, spurs and painted coverage.
Keep the existing allocator report in one ignored test and document the
current baseline without changing tracing behaviour or dependencies.
Co-Authored-By: Codex <noreply@openai.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A small closed loop whose fit collapsed to a single cubic (fit_closed),
or a piece returning to its own junction with fewer than two fitted
cubics (fit_piece), kept every positioned point as a straight segment.
A looser Paths collapses more loops, so it could give more segments
than a tighter one.
Such a loop now tries the fewest segments that still hold it, each only
if the result keeps an area (more than 0.25 px^2) and stays within
1.5 * max(tolerance, 0.5 px) of the points both ways: the fit at up to
three somewhat tighter tolerances (what a slightly higher Paths gives);
for a closed loop then two cubics split at its start and its farthest
point, then the outline simplified at the tolerance and finer. The
positioned points stay the last resort, as before.
Test (code-drawn): a_higher_paths_never_gives_fewer_segments sweeps
Paths 0/25/50/75/100 over 32 seeded 48 px images of soft blobs plus
pixel noise in Color and Black and White at Noise 1 and 3. Before, six
runs had fewer segments at a higher Paths (e.g. 125 -> 79 from Paths 0
to 25); after, none. One image (seed 16, Black and White: 140 -> 136
from Paths 0 to 25, with or without this change) has another cause and
is left out of the sweep.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018tBcdwNhptqtmX32JCGFzB
Where the pixel at a sharp corner falls to the other colour, the outline
cuts the corner in two one-pixel steps, and between two straight edges
of three cracks or more subpixel::outline makes those steps one point
(a notch). Whether the edges were long enough was measured on the whole
loop. An edge can run on past a junction in one region's loop while
the region across it, whose loop turns at that junction, sees a shorter
edge: one side made the notch and the other did not, so the shared
piece got different points on each side and the abutting curves left a
gap.
The notch now also requires the three cracks of each edge next to it to
lie in the piece (no junction among them), as the corner-pixel shift
correction above it already does.
Test (code-drawn): noisy_colour_images_leave_no_sample_uncovered traces
20 seeded 36 px images of soft overlapping blobs in four colours plus
pixel noise (Color, 5 colours, Noise 1 and 2, Paths 0/50/100) and needs
every one of five sample points per pixel to lie in some path: 93
uncovered before (seeds 9, 11, 13), 0 after.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018tBcdwNhptqtmX32JCGFzB
Layers::shift read a crack's sub-pixel shift from its inside pixel,
looking for a thin run on the inside before one on the outside. Where
both sides are thin (a one- or two-pixel sliver beside another, as in
soft gradients and noisy scans) the two regions of the crack got
different shifts, so their shared piece got different points and the
abutting curves left slivers that no path covers.
The shift is now the mean of the readings from both sides,
(s(c) - s(c flipped)) / 2, which is exactly antisymmetric in floating
point (Crack::flipped gives the crack seen from its other side).
Tests (code-drawn images):
- a_crack_between_thin_slivers_has_the_same_shift_from_either_side:
every crack's shift is the exact negative of its flipped crack's, on
a noisy ripple of 1-2 px columns at three thresholds.
- shared_boundaries::thin_slivers_on_both_sides_of_an_edge_leave_nothing_uncovered:
a 48 px noisy ripple in Black and White at thresholds 140/157/170 and
Paths 0/50/100, sampled on the shared-boundary check's 4x4 grid per
pixel: 482-1,228 uncovered points per trace before, 0 after.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018tBcdwNhptqtmX32JCGFzB
contour::trace_mask built a table of outgoing directions for every
pixel corner in a raster pass and then followed it. It now works out
each turn during the walk from the two pixels ahead of the current
corner: right if the one ahead on the right is outside, straight if the
one ahead on the left is, else left (so at a saddle it still turns
right and diagonal pixels stay apart). Loop starts are found by
scanning the pixels for a top edge on the outline, with one flag per
pixel for a top edge already walked. The walk is bounded by the grid
size, and a mask of the wrong length now gives no loops instead of
failing an assert_eq!.
The output is identical: the_walk_equals_the_stored_direction_walk
compares it with a test-only copy of the previous walk on all 65,536
4x4 masks and 400 seeded 23x17 masks.
Time (best of 7, release, 4 cores): the walk alone on 2000x2000 masks
19.1 vs 26.6 ms (smooth pattern), 81.3 vs 93.2 ms (12% random pixels),
223.1 vs 238.1 ms (50% random); whole 1000x1000 traces 73.1 vs 77.1 ms
(Black and White), 249.6 vs 251.7 ms (Color, 6 colours), 214.1 vs
218.6 ms (Color, 6 colours, Noise 1), same segment counts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018tBcdwNhptqtmX32JCGFzB
The 'Last updated · Change' status lines in docs/roadmap.md, docs/target-app-parity.md and docs/image-trace-quality.md change with almost every merge to main, so this PR's edits to them conflicted each time main moved. They are reverted to main's text; this PR's evidence bullets, history rows and table entries stay.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018tBcdwNhptqtmX32JCGFzB
The 'Last updated · Change' status lines in docs/roadmap.md, docs/target-app-parity.md and docs/image-trace-quality.md change with almost every merge to main, so this PR's edits to them conflicted each time main moved. They are reverted to main's text; this PR's evidence bullets, history rows and table entries stay.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018tBcdwNhptqtmX32JCGFzB
Gradients (off unless asked for) found a blurred or noisy linear gradient in
part or in pieces, with bands around it.
The areas that may be a gradient are tried largest first, and one that is a
gradient is grown over every pixel next to it whose colour is within 10 levels
of the gradient's there: the pixels noise put a little off, a soft rim, the
pieces noise cut off it.
Linear-gradient inputs with Gradients on, paths beside the gradient, before ->
after: blurred 3.0 -> 0.5 (6 Colors) and 55.5 -> 0.5 (High Fidelity Photo);
noisy 11.0 -> 0.5 and 23.5 -> 1.5; scanned 5.5 -> 3.0 and 46.5 -> 37.0; clean
and aliased as before; dE within 0.02. No gradient on any input that is flat
colour, with any preset (the harness fails if there is one).
Each documented preset has an enhanced version (`enhanced: true` in
`imageTrace.make`, an Enhanced checkbox under the panel's preset): the preset
with Auto Grouping, and with Shapes in all but High Fidelity Photo. Gradients
stay an option. Flat Logo, a preset of our own, has none. The legacy presets,
and documents traced with them, are as they were.
So that an enhanced preset is no worse than its preset on that preset's inputs:
- A hole that one colour fills is the same shape as the path that fills it, so
the two still share one outline (a shape beside a fitted hole left slivers).
- A rectangle is as large as where its sides lie on average, not their farthest
points.
- Auto Grouping keeps the paths of a group in the order they were traced in. It
put a path first and what is inside it on top: a white counter traced under
its black letter was then painted over the letter's edge, a light seam
(Default with Auto Grouping, edge dE 10.24 -> 9.88; legacy 10.05).
The harness measures every enhanced preset beside its preset.
Create: Strokes drew every line with round caps: a square-ended line 10 px wide
had its corners 2.94 px outside the stroke.
- A line's ends are square when its ink reaches as far a third of the width to
either side of its middle as along it (a round end stops short there by a
seventh of the width); told from lines 3 px wide and more, all the ends of a
line together. The stroke then has square caps (`TracedPath::square_caps`).
- The last points of a wide line's skeleton can turn towards a corner of its
end. An end that has left the line the skeleton followed a width back (by a
quarter of the width, 1 px at least) is put back on it; and an end may be
pulled back as well as pushed out to where the ink ends.
Outline on the thin-line inputs: largest distance from the art 2.94 -> 0.88 px
(clean 2.94 -> 0.41; aliased 0.88 -> 0.81), mean 0.055 -> 0.050 px, anchors
14 -> 10.
- A sharp corner may fall up to 0.75 px short within 3 px of where it should be:
anti-aliasing and blur take that much off a corner (not on a thin bar, which
is all corners).
- A rounded rectangle gets the radius that fits its outline best, searched
around the one its area says, instead of that one alone.
- A rectangle's direction is tried along its longest straight side, the longer
axis of its area and the pixel grid: a blurred one may have no straight side.
- A regular polygon's sides, corners and size are read from the outline's shape
(how its distance from the middle goes round, and its area), not from how many
anchors the fitter gave it.
- An outline without anti-aliasing is known to half a pixel: it gets half as
much room again.
Plain shapes found with Shapes on, all sizes and variants: Default 57% -> 98%,
6 Colors 62% -> 100%. The harness counts false shapes: none on any input but one
thin line (a rectangle 2 px wide) whose trace is 0.70 px off with or without
Shapes.
A scan's paper is a tenth darker in one corner than in another. Flat Logo found
each tone of it as a colour of its own and traced the bands and the specks
between them: on the scanned inputs 470 paths, 21 px from the art on average.
Two palette colours are now one when the pixels between them are no fewer than
a third of those at the lesser of the two: flat colours leave only their shared
edge between them, tones of one colour fill the gap. Whether the image is flat
art at all is still asked of the tones as found.
Flat Logo on its inputs, all sizes and variants: 3.106 -> 0.090 px off on
average, 88.5 -> 0.82 px at most, 69.9 -> 3.9 paths; clean, aliased, blurred
and noisy art as before.
Snap Curves To Lines made every straight piece within 2.5 degrees of level or
upright exactly so, whatever its length: a 200 px line 2 degrees off had each end
moved 3.5 px. Now a piece is levelled only when neither end moves more than
1.6 px; a longer line that far off is left as the image has it.
Technical Drawing (the preset with the option on) on its inputs: largest
distance from the art 4.21 -> 1.46 px (1.43 before levelling existed), mean
0.178 -> 0.126 px; on the askew inputs 98% -> 70% of the outline comes out level
or upright.
Gradients (params `gradients`, a checkbox in the panel for Grayscale and Color;
off unless asked for): an area whose colour changes evenly along one direction
is traced as one path filled with a linear gradient, not as bands of flat colour.
Gradients are found before the palette is made, so no colour is spent on their
bands: a connected run of smoothly changing pixels is a gradient when a plane
fits each colour channel with all three sloping the same way, the colours along
that way are at most six stops with ramps between them, and painting them back
gives the area's own pixels. Flat colour, soft edges between flat colours and
shading that goes two ways are left as they were.
The harness gains two linear-gradient inputs with known lines and stops, and
measures whether each is found, the angle of its line and the colours of its ramp.
Auto Grouping (params `autoGrouping`, a checkbox in the panel; off unless asked
for): each traced path that has others inside its outer contour comes out in a
group with them, the path first, nested as deep as the picture is. The path that
covers the whole image is nobody's group.
The grouped result paints the same picture: where paths can overlap (stacked
layers, Flat Logo, strokes) a path only joins a path below it, and only if that
moves it under nothing that shares any of its box.
The harness measures three presets with the option on ("grouped %"), painting
them in the grouped order.
Shapes (params `shapes`, a checkbox in the panel; off unless asked for): a filled
outline that is a circle, an ellipse, a rectangle or a regular polygon comes out
as that live shape. A shape is only made when every point of the traced outline
lies within 0.5 px (0.2% of the shape's size, if more) of it; an ellipse whose
radii differ by less than twice that is a circle of its area.
The outline before fitting: a corner's own pixel no longer pulls the two edges
that meet there, and a one-pixel step between two long edges is the corner where
they meet. Sharp corners kept: Default 69% -> 76%, the colour presets 86% -> 91%.
The harness gains plain-shape inputs and "shapes %".
Create: Strokes (the Outline preset, or the option) traced lines along the pixel skeleton: right to a pixel,
with a width counted in pixels, and fitted by a separate, older fitter.
- Each skeleton point is moved to the middle of the line's ink, read from the image across the line: every
pixel counts as much ink as its colour is the line's rather than the paper's, the ink's centroid is the
middle and its sum the width (an anti-aliased or blurred line says both to a fraction of a pixel).
- The stroke width is the median of those widths, no longer a whole number of pixels.
- Each end goes where the ink ends, less the half width a round cap adds.
- Centre lines are fitted as outlines are: a straight line is two anchors, corners stay corners, Snap Curves To
Lines levels them. The old Douglas-Peucker and simplify fitter, and the crate's pathops dependency, go.
Outline on the thin-line inputs, all sizes and variants (docs/image-trace-quality.md), before -> after:
- painted outline 0.310 -> 0.055 px off on average; 81 -> 14 anchors for five lines
- centre line 0.295 -> 0.052 px from the true centre line; stroke width 0.13 -> 0.02 px off; 18.4 -> 2.7
anchors a stroke (blurred, noisy and scanned lines: 2.0 each; aliased: 4.6)
- largest distance 1.84 -> 2.94 px: the square ends of a 10 px line under a round cap
Presets without Create: Strokes are unchanged.
The documented option "replaces slightly curved lines with straight lines and snaps lines near 0 or 90 degrees
to absolute 0 or 90 degrees" (helpx, Image Trace panel options); ours did the first half only.
- With Snap Curves To Lines, a straight piece at least 2 px long and within 2.5 degrees of level or upright is
made exactly so, by moving its ends across it to their mean (the page gives no angle: 2.5 degrees is our
choice). The pieces next to a moved end follow it.
- A junction other regions share stays where it is, so abutting regions still have one curve along their edge;
a line between a junction and a free end is levelled to the junction.
- The harness gains askew inputs (a rectangle, a plus sign and two lines, 1 to 2 degrees off level and upright)
and an "upright %" column: the share of the outline length that is exactly level or upright.
On the clean askew inputs (docs/image-trace-quality.md): Black and White Logo and Technical Drawing, the presets
with the option on, go from 0% to 91% and 95% of their outline exactly level or upright. Their outlines are then
0.49 and 0.59 px from the art as drawn where they were 0.05 (4.2 px at the far end of a 300 px side turned 1.5
degrees): that distance is what the option is for. Presets without the option are unchanged.
Against the documented panel (https://helpx.adobe.com/illustrator/desktop/manage-objects/traces-mockups-symbols/image-trace-panel-options.html):
- Ignore Color: "Ignores a specific color while tracing. Use the color picker to selectively ignore a color."
The panel's Ignore White is now Ignore Color with a colour button beside it (white until another colour is
picked); `imageTrace.make` takes `ignoreColor: "#rrggbb"` beside `ignoreWhite`, and leaves out the palette
colours within about 30 levels of it. Documents saved before have no such setting and ignore white as they
did. Flat Logo mode leaves out white only.
- Auto Color ("Creates a posterized image from a photo or artwork"): Color mode with the Automatic palette.
- Outline ("Simplifies the image to black outlines"): Black and White with Create: Strokes and no fills.
- The Outline preset's name is translated as a preset's name, not as the Outline view's.
Measured (docs/image-trace-quality.md): Auto Color on flat art is 0.140 px off on average with 4.0 paths (16
Colors: 0.141, 4.2). Outline on the line inputs, the first measure of Create: Strokes: 0.310 px off on average,
1.84 at worst, 81 anchors for 5 lines (Line Art's filled outlines: 0.055, 0.96, 46).
Black and White read every pixel as a blend of "black" and "white", and took for those the mean luminance of
the image's black and white pixels. In blurred or scanned line art most black pixels are part paper, so the
mean black was a gray: the middle of every line counted as more than all ink, and by how much changed as the
line crossed the pixel rows, which put a ripple on its edges that the fit then followed with anchors.
- Black is the image's darkest black (the 2nd percentile of its black pixels' luminance).
- White is the lightest pixel within a few pixels of each crack: the paper as it is there. A scan's paper is
darker in one corner than in another, and read against the image's lightest white it counted as part ink.
On each preset's own inputs, all sizes and variants (docs/image-trace-quality.md), before -> after:
- Line Art: 78 -> 46 anchors, outlines 0.093 -> 0.055 px off on average (blurred lines 168 -> 53 anchors,
scanned 199 -> 70); before sub-pixel edges it was 25 anchors and 0.158 px off
- Technical Drawing: 56 -> 46 anchors, 0.092 -> 0.084 px off
- Sketched Art: 0.197 -> 0.186 px off; Default 0.092 -> 0.090; Black and White Logo 0.063 -> 0.059 with
27 -> 29 anchors
Not kept: reading lines up to six pixels deep by their ink's centroid and width, and a looser tolerance for
straight runs, neither of which measured better.
Every colour layer took a mask, component labels, a fill stack and an edge map the size of the whole image,
and as many layers were traced at once as the machine has cores.
- A layer is traced within the box its pixels lie in.
- The stack that fills components holds `u32` indices (a mask has at most 64 million pixels).
- Layers are traced as many at a time as fit in half a gigabyte (about ten bytes a pixel of the largest
layer's box each), on however many cores.
Peak memory above the image itself, photo-like image (docs/image-trace-quality.md), before -> after:
- 16 megapixels: High Fidelity Photo 3,856 -> 880 MB (2.85 -> 3.65 s), Low Fidelity Photo 1,972 -> 754 MB
(1.49 -> 1.80 s), 16 Colors 1,625 -> 298 MB, Shades of Gray 852 -> 407 MB, Default 256 -> 199 MB
- 4 megapixels: High Fidelity Photo 981 -> 643 MB, 16 Colors 408 -> 122 MB, times unchanged
The traces are the same, path for path.
Each region's outline was fitted on its own, so the edge between two regions was fitted twice, a little
differently, and left hairline gaps and overlaps.
- An outline is cut at its junctions: the pixel corners where three or more regions meet, where two parts of a
region touch by a corner only, and where an edge reaches the image's border (each side of the image is a
neighbour of its own, so the image's corners stay exact). A junction lies where the edges around it say, the
same for every region around it.
- Each piece between two junctions is the whole edge between two regions. Both get exactly the same points for
it (shifts are rounded to 1/1024 px so that the two sides agree to the last bit; whether a piece is hard is
decided by the piece alone) and it is fitted the same way from either side (from its end further left), so it
is one and the same curve in both paths. A loop without junctions is fitted clockwise from its leftmost point:
a shape's outline and the hole it leaves in the shape around it are one curve too.
- Stacked layers (Overlapping) share no edges and are cut at the image's border only.
On each preset's own inputs, all sizes and variants (docs/image-trace-quality.md), before -> after:
- Default: gaps 22.2 -> 0.8 px, overlaps 22.4 -> 0.5 px
- 3 Colors: gaps 48.6 -> 0.0, overlaps 29.4 -> 0.0; outlines 0.087 -> 0.104 px off on average
- 6 Colors: gaps 39.0 -> 0.2, overlaps 35.7 -> 0.3; 0.137 -> 0.140 px off, largest 1.09 -> 1.34
- 16 Colors: gaps 34.5 -> 0.2, overlaps 26.4 -> 0.2; 0.133 -> 0.141 px off, largest 1.09 -> 1.31
A photo's noise is as large as the step from one palette colour to the next, so noise decided pixel by pixel
which of two neighbouring colours a pixel got: tens of thousands of specks, or (since the palette change)
neighbouring tones merged into one.
- A continuous-tone image (eight colours leave much of it unsaid; flat art, however noisy or unevenly lit, is
not one) is smoothed before it is quantised: three rounds of a small bilateral filter, which averages noise
away within a tone and keeps the edges between tones where and as sharp as they were.
- With Method: Overlapping, layers are stacked by lightness, from the largest colour's end of the scale,
instead of by area. Each layer reaches under those above it, so where two colours meet every layer between
them in the stack ends on that same edge and shows through its anti-aliasing: by area that was an unrelated
colour's hairline around every region; by lightness it is a neighbouring tone.
- Ignore White with Overlapping: no layer reaches under the white that is left out.
On each preset's own inputs, all sizes and variants (docs/image-trace-quality.md), before -> after:
- High Fidelity Photo: colour error 2.37 -> 1.75 (2.30 at the baseline), edge error 2.59 -> 1.98, 639 -> 121
paths (3,653 at the baseline), 25,604 -> 9,818 anchors, specks 152 -> 10; a 16 megapixel photo-like image is
traced in 3.5 s (31,131 paths) where it was refused
- Low Fidelity Photo: colour error 2.67 -> 2.45, 55 -> 37 paths
- Shades of Gray: 73 -> 17 paths, colour error 9.33 -> 9.26
- The other presets' numbers are unchanged
Color and Grayscale chose their colours from every pixel, so the blends along anti-aliased edges took palette
entries and traced as slivers, and with more colours asked for than the art has, k-means split flat colours
along their noise and the halves traced as specks all over each other.
- The palette is made from the picture's own colours: a pixel that lies between the pixels on opposite sides
of it is a blend and counts for nothing (lines one pixel wide and noise are no blends).
- Colours that noise made of one colour merge: close colours whose pixels are mixed through each other. The
bands of a gradient keep to themselves and stay apart.
- A blend pixel whose nearest palette colour is a third colour (halfway from navy to yellow lies an olive that
is nearer to a green elsewhere) goes to the nearer of the two colours it lies between.
- Boxes are cut where their two sides differ most (Otsu's criterion) instead of at the median, which split the
largest colour of flat art along its noise and left two smaller colours in one box.
- Colours are compared in Oklab when pixels are assigned and in k-means.
On each preset's own inputs, all sizes and variants (docs/image-trace-quality.md), before -> after:
- 16 Colors: 69.9 -> 4.2 paths, 1743 -> 66 anchors, outlines 5.08 -> 0.13 px off on average (largest 85.9 ->
1.09), kinks 1376 -> 4.4, paths in a colour the art doesn't have 5.9 -> 0
- 6 Colors: 21.2 -> 4.0 paths, 4.44 -> 0.14 px off; 3 Colors: colour error 4.86 -> 0.43, 0.33 -> 0.09 px off
- High Fidelity Photo: 3636 -> 639 paths, 81,597 -> 25,604 anchors, colour error 2.50 -> 2.37; a 1 megapixel
photo-like image 1.98 -> 0.42 s (51,389 -> 4,298 paths), a 4 megapixel one traced (27,583 paths) where it was
refused
- Low Fidelity Photo: 128 -> 55 paths, colour error 2.59 -> 2.67; Shades of Gray: 91 -> 73 paths, 9.29 -> 9.33
Each crack between an inside and an outside pixel is moved to where the pixels across it say the edge is (the
blend of the two tones in the column of pixels across the crack; a thin line by the centroid and sum of its
tone; hard-edged images are read as staircases), and the outline is fitted as Flat Logo mode fits its own:
corners put back, straight runs as lines, a round outline as a circle, the rest as few curves as Paths allows.
Snap Curves To Lines straightens slightly curved runs and makes arcs. A fit that strays from its outline
gives way to a plainer one, and at last to the outline's own points.
On each preset's own inputs, all sizes and variants (docs/image-trace-quality.md):
- Default: outlines 0.313 -> 0.091 px off on average, largest 1.51 -> 1.02; edge error 17.2 -> 10.1;
gaps between abutting paths 63.5 -> 22.2 px, overlaps 67.0 -> 22.4 px
- Black and White Logo: 0.270 -> 0.064 px off, kinks 15.3 -> 0.3, anchors 3.78 -> 2.73 times the art's
- Technical Drawing: 0.201 -> 0.086 px off; Line Art: 0.158 -> 0.093
- Worse: anchors on blurred and scanned thin lines (Line Art 25 -> 77 on average), High Fidelity Photo
(colour error 2.30 -> 2.50, anchors 46,866 -> 81,597, 1 megapixel photo-like image 0.37 -> 1.98 s)
crates/trace/tests/quality draws its inputs from known vector art (no image files), traces them with every
built-in preset and measures fidelity, distance from the true outlines, economy, artefacts, time and memory.
docs/image-trace-quality.md records the numbers on main before any change, beside the same measures of an
open tracer run as a black box. No change to the tracer.
