From raster image to candidate set
The browser first decodes and analyzes the source image. It looks at dimensions, transparency, edge and color characteristics, likely background, and whether the artwork behaves more like monochrome or multicolor input. The uploaded file remains unchanged and conversion stays on the device.
That analysis combines with the selected use case to build a candidate set. A typical automatic run evaluates about 16 tracing configurations, although the exact count can vary by image and target. Images with one color or a few very similar colors take a monochrome-oriented route while retaining the detected foreground color for the result swatch.
For flat-color Web and Logo artwork, the search can also include 2-, 3-, 4-, 6-, and 8-color candidates. Those passes give a small, coherent palette a chance to win against a trace that gains only a little visual similarity by creating many extra shades and paths.
What changes between passes
Candidates explore meaningful combinations of tracing behavior, such as color handling, region clustering, curve and detail simplification, and eligible source enhancement. VectorizeThis does not blindly generate every possible combination. The source profile and use case narrow the search to configurations likely to be useful.
The precise weights, thresholds, and candidate recipes are internal implementation details. Keeping them private lets the product evolve without promising that one numeric formula will remain permanent.
How candidates are compared
Each candidate is rendered and measured against the source. Selection considers several kinds of evidence:
- Visual resemblance, including structural and color similarity.
- Geometry, including silhouette, visible components, holes, and closed regions where the use case requires them.
- Complexity, including file size, paths, commands, nodes, segments, subpaths, and small isolated regions.
- Target eligibility, such as duplicate or open geometry checks for fabrication-oriented output.
These signals are interpreted together. The SVG with the highest visual-similarity measurement is not automatically the best product result if it is dramatically larger, noisy, or unsuitable for the chosen target.
Balanced winner and SSIM winner
The default winner is the strongest eligible balance for the selected use case. SSIM is one of the visual-similarity measurements. SSIM regret reports the difference between the balanced winner and the candidate with the highest measured SSIM.
A regret of zero means the balanced result is also the measured SSIM winner. A small positive value means VectorizeThis accepted a limited similarity difference for another advantage, often simpler or safer geometry. Desktop users can reveal the technical diagnostics through Output details and apply Use SSIM Winner for comparison.
Final SVG precision comparison
After a tracing candidate wins, VectorizeThis tests optimized SVG finalists at decimal precisions 0, 1, and 2. It keeps detailed coordinates only when their measured benefit justifies the extra bytes. The manual SVG precision setting also offers 3 for cases where you want to inspect an even finer output.
Output preparation and checks
The winning trace is then prepared for its target. This can include SVG optimization, responsive viewBox handling, background policy, optional cropping, color exposure, and fabrication-oriented geometry checks. Output preparation happens after candidate scoring so minification alone does not make a trace look like a better visual candidate.
Warnings describe measurable concerns; they are not machine certification. Tiny-island cleanup is offered only when relevant, and fabrication errors can block download until they are resolved.
What happens after a manual change
When you change a setting on a result card, VectorizeThis recalculates that image with the selected value and updates the same card. It does not need to replay the complete initial search for every adjustment. Color edits are reapplied when the regenerated SVG has the same number of colors. Bulk edit uses these same per-card controls and skips cards that already match the requested value.
Practical limits
Automatic selection compares the candidates it generates; it cannot recover detail that is absent from the raster or recreate the decisions in an original vector source. Compression artifacts, antialiasing, shadows, gradients, low resolution, and tiny features can all affect the result. Fabrication also depends on software, machines, material, tooling, and physical tolerances that an image trace cannot fully validate.
Start with the automatic winner, compare both previews, and use the settings reference only where you can identify a specific problem. For a broader product overview, read How VectorizeThis Works.
Last updated: