Image to Vector Conversion: How It Actually Works

Two Fundamentally Different Ways to Store an Image

A raster image is a grid of coloured squares. A 3000 x 2000 JPEG is six million pixels, each with its own colour. Enlarge it and you are enlarging squares, which is why zoomed photos go blocky. The file has no concept of a "line" or an "edge" — only neighbouring pixels that happen to differ.

A vector image is a set of instructions: start at this point, curve to that one with these control points, fill the enclosed area with this colour. There are no pixels, so there is no resolution. The same file renders sharp on a phone screen and on a billboard.

Conversion means going from the first to the second, and that is a genuinely hard problem — you are asking software to look at a grid of colours and infer the shapes a human would perceive. There is no single correct answer, which is precisely why the results vary.

Cutting and engraving machines need the second kind. A laser head follows a path; it cannot follow a pixel. This is covered in more depth in SVG vs PNG for laser.

The Four Stages of a Conversion

Almost every vectoriser, ours included, runs some version of these four stages. Knowing which stage failed tells you which control to reach for.

1. Separation. Decide what is subject and what is background. Traditional tools skip this entirely and treat the whole frame equally, which is why backgrounds bleed into results. Modern tools use a segmentation model to isolate the subject first.
2. Reduction. Collapse millions of colours into a small number of regions — often just black and white. This is where a threshold is applied, and where most information is lost. It is also where most failures originate.
3. Cleanup. Remove speckle, close small gaps, thin thick blobs down to consistent strokes. Without this stage, tracing faithfully reproduces every bit of sensor noise as a tiny path.
4. Tracing. Fit mathematical curves to the boundaries of the remaining regions. The tracer decides how closely to follow every wobble — too closely and you get thousands of nodes, too loosely and detail rounds away.

When a conversion disappoints, identify the stage: background contamination is stage one, a blobby or washed-out result is stage two, speckle is stage three, and jagged or over-smoothed curves are stage four.

Why the Same Settings Give Different Results

Threshold is an absolute brightness cutoff, but photographs are not exposed consistently. A threshold of 50% applied to a photo shot in bright sun keeps almost everything; the same 50% on an underexposed indoor shot turns the whole frame black.

Three properties of the source dominate the outcome:

• Contrast range. An image using only the middle of the brightness range has no natural cutoff point — every threshold value is a bad compromise. Increasing contrast before conversion helps far more than any setting afterwards.
• Noise level. High-ISO images have per-pixel variation that the tracer faithfully reproduces as speckle. Denoising before conversion is usually the fix.
• Edge softness. Shallow depth of field and motion blur mean edges span many pixels rather than one. Any cutoff placed inside that transition is arbitrary, and the resulting outline wanders.

This is why "what settings should I use" has no useful universal answer, and why five minutes spent improving the input beats an hour of adjusting sliders. Choosing the right photo covers the input side in detail.

Reading a Bad Result

Preparing an Image So It Converts Well

Five minutes in any photo editor before conversion is the highest-return work available:

1. Crop tight. Everything outside the subject is a source of unwanted paths.
2. Increase contrast. Push the histogram towards both ends. A clear separation between light and dark gives the threshold something real to work with.
3. Denoise. Especially on phone photos taken indoors. Noise becomes speckle, and speckle becomes hundreds of useless paths.
4. Convert to greyscale mentally. Two colours of similar brightness — red on green, for instance — become identical grey and merge. If the subject only separates from its background by hue, adjust before converting.
5. Do not upscale. Enlarging a small image invents pixels that were never measured, and the tracer will trace the invention. Start with the largest original you have.

Resolution guidance: roughly 1000–2000 pixels across the subject is the sweet spot. Below that there is not enough information to find clean edges. Far above it, you are mostly tracing noise at very high precision.

What Vector Conversion Cannot Do

Worth stating plainly, because a lot of frustration comes from expecting the impossible:

• It cannot recover detail that was never captured. A blurry photo produces a blurry-edged vector. Vectorising does not sharpen.
• It cannot make a low-resolution image high-resolution. The output scales infinitely, but it scales whatever shapes were found in the original pixels — including their inaccuracies.
• It cannot reproduce continuous tone as vector fill. Gradients become bands. This is a mathematical property of the format, not a limitation of any particular tool.
• It cannot decide what matters. A person knows the eyes matter more than the wallpaper. Automatic tools weigh both by contrast alone, which is why manual cleanup still improves nearly every result.