Changelog
All notable changes. Benchmarks for every engine version are in docs/bench.md.
1.7.0 — 2026-11-02 — measured accuracy
Found with a new benchmark of 45 license-free pictures and a visible-defect metric; every number below compares 1.6.0 and this version on the same picture (docs/research.md §7).
- Fix: the Logo style no longer punches holes into opaque pictures. The background color was a candidate for every pixel even when no background had been found, so art of the border's color turned transparent: a dark 1713×918 painting 80.2 % → 94.4 % similarity, visible defects 59.6 % → 33.0 %. A background guessed from the border is kept when the picture turns out not to be flat (with a note). A flag's white band is a shape, not a hole.
- Logo style keeps every flat color: the automatic palette also requires 99.8 % of the flat pixels to match (up to 16 colors), so a color and its darker shade stay two colors. Flat shop icon: 97.99 % → 99.95 %, visible defects 18.33 % → 0.00 %.
- Fix: a flat color between two others (grey between black and white) is no longer taken for anti-aliasing: Tux icon 97.65 % → 99.77 %, visible defects 10.22 % → 0.01 %.
- Fix: thin straight bars (4–12 px) keep their width instead of becoming zero-area slivers: 94.92 % → 100 %.
- Pixel art is exact: every color (up to 64) and every single pixel, also in checkerboards. A sprite sheet: 97.58 % → 100.000 %, visible defects 16.26 % → 0 %.
- Logo style keeps small details: pieces down to 0.003 % of the artwork (was 0.05 %), so a flag's stars and motto stay (found by the vision judge, which scored the flag 4/10).
- Logo style: light halos left by an older trace return to the background. The "not a flat logo" note is right more often: line art no longer gets it; shading and gradients cut into bands do.
- Small pictures (under 512 px) are segmented on a 2× grid by the Cartoon, Flat and Poster styles (
prescaleoption; coordinates and curve tolerances stay in source pixels): text and 1–2 px lines survive. A 235 px icon with text: 98.45 % → 99.03 % without AI. - Cartoon style: every ink stroke keeps its color (up to three inks, so grey lines stay grey next to black ones), and a dark area wider than a stroke with shading of its own (a reflection on a black screen) stays a fill instead of turning black. On 32 Cartoon / Flat / Poster pictures: 6 better, none worse, median visible defects 0.49 % → 0.31 %; futuristic mascot 5.63 % → 0.03 %, fish mascot 1.66 % → 0.17 %, robot 1.24 % → 0.23 %.
- Logo style, from the literature (docs/research.md §8): rings and lines are no longer cut into pieces at every thin spot (a line drawing: 31 → 9 paths, 6.7 → 2.2 KB, 98.90 % → 99.62 %); ellipses and circular arcs become exact curves (direct least squares, Fitzgibbon et al. 1999); a run is a line only if its residual is not a systematic bow.
- Logo style with the background kept (an icon on an opaque page): the page no longer crowds the art's colors out of the palette. Shop icon on white: 97.75 % → 99.96 %, visible defects 10.42 % → 0.00 %.
- Logo style on noisy pictures (scans, photos of prints): the noise is estimated from the picture and smoothed away with an edge-preserving filter first; clean pictures are untouched. A flag under strong noise: 83.0 % → 98.5 %, visible defects 36.7 % → 0.00 %.
- Automatic AI mode: with the Logo style only for JPEGs (on small PNG logos it did not help; on JPEG logos it cuts the visible defects from a median 0.53 % to 0.02 % on 12 compressed logos), never with Pixel art.
pngsvg.compare:defect_area/defect_map(visible defects, CIEDE2000 > 5 without anti-aliasing lines),psnr,ssim,edge_split;rendertakes a background (a checkerboard shows holes).- Benchmark:
bench/readme.py(both sets, the style per kind of picture, small-picture study with and without AI),bench/evidence.py(pictures of the results),bench/before_after.py(1.6.0 vs now),bench/judge.py(optional vision-model review via NVIDIA NIM). New set E8 of 18 pictures made as rasters (fetch_sets.py --set real).bench/compare_tools.py(VTracer and Potrace on the same pictures),bench/versions.py+bench/stats.py(paired statistics: Wilcoxon signed-rank, bootstrap confidence intervals, effect sizes) andbench/robustness.py(JPEG, down-scaling and noise, scored against the clean picture).
1.6.0 — 2026-10-25 — logo engine
- Logo / icon now uses a new flat engine (docs/logo-engine.md): one
<path>per shape with a stableidand colorclass, straight edges and sharp corners (lines snapped to 0°/90° and parallel sides aligned), shared boundaries between neighbours (nostrokegap filler), transparent background (no<rect>), anti-aliasing absorbed instead of traced, shapes touching through a thin neck split,viewBoxtrimmed to the artwork. The Turion logo: 4 paths in 468 bytes (before: 5 paths, 3.9 KB with the logo style; 7 paths, 19.7 KB with the cartoon style), IoU ≥ 0.997 per color. - New Pixel art style: edges follow the pixel grid.
- Logo options in Fine-tune, profiles and the CLI: corner rounding (
--corner-radius), sharp-corner angle, straighten angles (--no-snap), split touching shapes (--no-split), trim (--no-trim); the fine-tune panel only shows the options of the chosen style's engine. Results report where a trimmed SVG sits (stats.crop) and a warning when a picture is not flat artwork (stats.warnings). pngsvg.color.delta_e2000(CIEDE2000, checked against Sharma et al.'s reference pairs).- License: GNU AGPL-3.0 (was MIT): copies, modified versions and web services built on it must stay open source and keep the attribution to the author (additional terms under section 7 in NOTICE). Releases 1.0.0–1.5.0 remain available under MIT.
- Python 3.11 or newer is required for the package and command line (3.10 reached end of life in October 2026), so every install gets Pillow ≥ 13.0 with its security fixes for image decoding. The desktop app bundles Python 3.13.
Website — 2026-10-25
- Product website at https://ridhuandev.github.io/png-to-svg-vectorize/ (
site/, Next.js static export on GitHub Pages): features, how it works, gallery with measured numbers, honest comparison, downloads with SHA-256, guide and release notes from this repository, in English and Bahasa Indonesia. No trackers. - Releases now include
SHA256SUMS.txt.
1.5.0 — 2026-10-25 — center lines, editing, automation
- Center lines (Shapes → Center lines, CLI
--centerline): line drawings become one stroke along the middle of each line, without fills — for plotters, laser and embroidery. Zhang-Suen thinning, a junction graph with spur pruning, and the existing curve fitting; stroke width measured from the ink. PDF/EPS/DXF export the strokes (DXF: open polylines on layerCENTERLINE). - Layers per color for design tools (
layers, CLI--layers): fills grouped in<g id="color-RRGGBB">. - Edit shapes (Compare → Edit, or
E): click a shape to select it, recolor it, give it another shape's color (merge), or delete it; undo / redo (Ctrl+Z/Ctrl+Shift+Z). Edits are saved to the service (PUT /api/jobs/{id}/svg), which accepts only the engine's own SVG subset (no scripts, links, entities or foreign markup; same viewBox), and all exports use the edited version. - Watch folder (desktop: Settings → Automation; CLI
png2svg watch IN OUT [-f pdf] [--profile P]): pictures that appear in a folder are converted with the current style into another folder, once their copy is complete. - Batch report: CSV or JSON of every result (file, style, size, time, colors, similarity — measured for results not opened yet). CSV cells that a spreadsheet would run as a formula are neutralized.
- Keyboard shortcuts with a sheet (
?): open, process, save, save all, previous/next result, remove, edit, undo/redo. - Benchmark set: Wikimedia re-encodes and sometimes re-renders its thumbnails;
bench/fetch_sets.pynow also accepts identical pixels (pixels_sha256) and known alternate renders (also_sha256). - Release builds use Pillow 13.0.0 (security fixes for image decoding; owner-approved exception to the two-week rule for dependency age). Benchmark output unchanged on all 27 E7 images.
1.4.0 — 2026-10-25 — studio options
- Colors: Number of colors (2–32 flat colors, ink included) or a locked palette: Use these colors on a
result, then change, add or remove colors (a removed color merges into the nearest); Grayscale and
Black & white (with threshold). Results report the colors used (
stats.palette). - Transparent background: transparent pixels, or else the region covering most of the border, are left out. Transparency survives the AI step.
- Shapes: stacked or cut out — cut-out shapes keep their holes and never overlap (cutting machines, editing); gap filler (seam width) and speck removal (minimum area) are now fine-tune options.
- Export formats: SVG, PDF (vector, real gradients), EPS (level 3), DXF (R12 polylines, one layer per color) for the selected result or Save all; PNG rendered from the vector at 1–8×; Copy PNG.
- Paste a picture with
Ctrl+V. - Style profiles: save the current style under a name, pick it again, export/import
<name>.pngsvg-profile.json; the command line reads the same file (png2svg --profile). - CLI:
--format,--max-colors,--palette,--color-mode,--threshold,--remove-bg,--cutout,--seam,--speckle,--profile;-o out.pdfpicks the format from the extension. - API:
GET /api/jobs/{id}/export?format=,GET /api/export.zip?format=,POST /api/profiles/check; new job optionsmax_colors,palette,color_mode,threshold,remove_bg,layout,seam,speckle. - Default output is unchanged (byte-identical); Vite 8.3.4 (security fixes in the dev server).
1.3.0 — 2026-10-25 — Enhance with AI (optional)
- Enhance with AI: before vectorizing, an AI model (Real-ESRGAN, BSD-3-Clause, converted to ONNX) cleans and enlarges the photo ×4; the result is traced at 2× the photo size. Biggest effect on small, blurry and JPEG pictures (+0.47 similarity points median on damaged copies of E7 images, see docs/bench.md).
- Off / Automatic (default: only photos under 1000 px or JPEGs) / Always — in the style panel. Two models: Illustrations (default, 17 MB) and General (5 MB).
- Models are not bundled: Settings → Processing → Download, once, from this repository's GitHub release
models-v1, verified by SHA-256 and kept in the user data folder; works offline afterwards. Cancel/remove at any time. The ONNX Runtime ships in the installer (DirectML on Windows — any GPU — and CoreML on Apple Silicon, with CPU fallback); Run AI on: Automatic / Graphics card / Processor only. Not available on macOS Intel (no runtime build). - API:
GET /api/ai,POST/DELETE /api/ai/models/{name},POST /api/ai/models/{name}/cancel; job optionsai(model name) andai_auto;PUT /api/settingsacceptsai_accel; results reportstats.ai. An AI job whose model is missing is refused with 422 instead of silently running without AI. - Maintainers:
tools/models/convert.pyconverts the official weights (loaded withweights_only=True) and checks the ONNX output against PyTorch;bench/ai_restore.pymeasures the gain. Python extra[ai].
1.2.0 — 2026-10-18 — themes and Settings
- Three themes to choose from — Paper & ink, Lightbox, Blueprint — or Match system (default). Colors live
in
app/ui/src/theme/themes.json;npm run checkverifies WCAG contrast for every theme and that components use no literal colors. - Bundled IBM Plex Sans / Mono (SIL OFL); numbers use the monospaced face. Pen-node motif in titles and the new brand mark.
- Settings (gear or
Ctrl+,): language, theme, density, text size, reduce animation, guide; processing speed, largest photo, workbench size; style for new photos, save folder (desktop), file name pattern ({name}{style}{date}), coordinate precision (engine optiondecimals, 0–3), width/height in the SVG; diagnostic log (desktop); version, licenses and a manual update check (GET /api/update). - Languages from files:
app/ui/src/locales/<code>.json; a new language is one file (see CONTRIBUTING.md). - API:
PUT /api/settingsacceptsname_pattern,size_attrs,max_side,max_staged(each optional); jobs carryfilename; saved/zip/downloaded SVGs follow the pattern and size setting.
1.1.0 — 2026-10-13 — friendlier workflow
- Workbench: added photos wait (status
staged) until you press Process N photos; tick only the ones you want. API:POST /api/jobs?start=false,POST /api/jobs/start(ids + style),POST /api/queue/cancel. - Queue that spares the computer: a real FIFO queue with a limit on simultaneous conversions; processing speed
Automatic (default, from cores and memory) / Saver (1) / Balanced (2) / Fast (≤ 4, one core left free), changeable
at any time (
GET/PUT /api/settings); workers run at below-normal priority; queue position shown per photo. - New interface: custom components instead of browser/OS controls (buttons, checkbox, switch, segmented control, dropdown, slider, progress ring, dialog, toasts with Undo), step indicator, result styles as visual cards, micro-interactions (press, hover, drop zone, finish animation) that respect reduce motion, plain-language messages for every failure, keyboard and screen-reader support.
- First-run guide (skippable, Esc; reopen with Guide).
- Removing a result can be undone for a few seconds; Stop all asks first; closing the desktop window while photos are still being converted asks first (in the page language).
- Desktop "Open" now puts photos on the workbench instead of converting them at once.
1.0.0 — 2026-10-12
First release of the desktop app for Windows and macOS.
- User guide in English and Bahasa Indonesia (docs/user-guide.en.md, docs/user-guide.id.md), screenshots, release notes.
- Comparison set of 27 CC0 / public-domain images (
bench/sets, downloaded on demand and verified by SHA-256), checked in CI together with the synthetic set; template for manual comparisons with other tools. - Fix: transparent PNGs are flattened onto white instead of black.
- Fix: when the darkest pixels are a colored dark area (a purple flag, deep water) rather than lines, the outline color is taken from line pixels.
bench/run.py --preset/--only.- Fix:
python -m pngsvgwith several files and worker processes failed on Windows/macOS (the worker function lived in__main__); the CLI now lives inpngsvg.cli.
0.5.0 — 2026-10-12 — desktop app and installers
- Desktop app (
pngsvg-app): native window (WebView2 / WKWebView) around the local service, native open / save / save-all dialogs (paths never go over HTTP, existing files are never overwritten), settings remembered between runs, log file in the user data folder,--smokeself-test. - Packaging: PyInstaller (onedir, no UPX), per-user Inno Setup installer with a WebView2 check, portable zip, DMG for
Apple Silicon and Intel.
Releaseworkflow builds and smoke-tests all three, draft release forv*tags. - Builds are not code-signed (see the user guide for the SmartScreen / Gatekeeper steps).
0.4.0 — 2026-10-12 — local web UI
png2svg serve: FastAPI service on127.0.0.1with a per-session token, Host check, CSP, upload limits (50 MB, 8000 px), content-based format check, validated options; jobs in a process pool with per-stage progress, cancel, re-run, delete, zip export. Results live in memory only.- UI (Svelte 5): drag and drop, queue, side-by-side and slider comparison with synchronized zoom up to 800 %, fills-only / outlines-only views, in-page similarity score, presets and advanced settings, save / save all / copy, English and Bahasa Indonesia, light and dark theme.
- Engine:
vectorize(..., on_stage=)progress callback (output unchanged).
0.3.0 — 2026-10-12 — radial gradients, speed, presets
- Elliptical radial gradients for blob shading; concentric bands of a ball or glow are grown back into one region.
- Gradient stops solved by least squares against what a renderer draws.
- Sub-pixel fill boundaries from anti-aliasing and boundary refinement (straight edges, sharp tips).
- Adaptive ink detection for thin outlines; wider anti-aliasing band merging.
- Segmentation about 2× faster; optional process pool (
executor=, CLI--jobs). - Presets
cartoon,flat,logo,poster(--preset).
0.2.0 — 2026-10-12 — shared boundaries and own curve fitting
- Planar map: every boundary between two regions is fitted once and shared (no gaps, no overlap).
- Schneider cubic Bezier fitting with corner detection; looser tolerance for boundaries hidden under outlines.
- vtracer dependency removed. Median file size −58 %, segments −69 % on the reference set.
0.1.0 — 2026-10-12 — first version
- Engine
pngsvg: ink outline separation by color unmixing, gradient-fill regions, pure-path SVG. - CLI
png2svg, tests, CI on Windows, macOS and Linux; benchmark scripts and baseline.