Why AI images vectorize badly, and how to prompt around it

An auto-tracer works by finding boundaries between regions of similar colour, then drawing a path along each boundary. It needs those boundaries to be abrupt. A hard transition from black to white is unambiguous: the path goes right there.

Generated images rarely give you that. Even when you ask for a flat illustration, the output usually carries a photographic finish: slightly soft edges, fills that vary by a few percent across a region, sometimes a hint of grain.

Your eye reads that as "clean flat illustration". The tracer reads it as a lot of small colour changes, and draws accordingly.

Anti-aliasing. Every edge has a band of intermediate pixels. The tracer either includes them, producing a doubled or fuzzy outline, or excludes them, eating into the shape. Neither is what you drew.

Gradients. There is no clean way to trace a gradient into a region-based vector. The tracer approximates with stacked bands, and a sky that looked smooth becomes visible stripes.

Texture and grain. Many models add a subtle paper or film texture, especially for anything described as illustration, poster, or vintage. That texture is noise at pixel level, and noise traces as speckle.

Near-identical colours. Generated flat fills are rarely one value. A region that reads as a single blue is often several blues within a few percent of each other. Depending on your colour-count setting, the tracer either merges them, losing detail elsewhere, or splits them, giving you four shapes where you wanted one.

These are working heuristics, not benchmark results. They're the terms that tend to help across ChatGPT, Midjourney, Flux and Ideogram, and they're worth testing against your own output:

• —flat vector illustration
• —solid fills, no gradients
• —bold thick outlines
• —high contrast
• —white background or transparent background where the model supports it
• —two colors or three colors, since naming a number tends to work better than saying "limited palette"
• —clean sharp edges
• —no texture, no grain, no noise
• —no shadow

Terms to drop if you plan to trace: 3d, render, cinematic, realistic lighting, depth of field, bokeh, watercolor, oil painting, grainy, film, vintage poster. Each one reintroduces exactly what you're trying to avoid.

Models tuned toward graphic output generally give more traceable results than models tuned toward photorealism, and telling a model explicitly that the output will be vectorized tends to help. Beyond that, be careful with specific recommendations, including mine: model behaviour shifts with versions, parameters, and subject matter, so a comparison that was true last quarter may not be true now. Test with your own prompts rather than trusting anyone's ranking.

If you want vector output without the trace step at all, tools like Recraft generate SVG natively. That's a different trade-off: less stylistic range, no conversion loss. Worth knowing about rather than pretending it doesn't exist.

Three cleanup steps, each a few seconds, each usually worth more than fiddling with tracer settings:

Posterize. Reduce the image to a small number of colours before tracing. This collapses near-identical fills into actual flat regions and removes much of the anti-aliasing ambiguity.

Raise contrast and clip the levels. Push the lightest tones to white and the darkest to black. The tracer then has something unambiguous to work with.

Handle the background deliberately. A clean transparent background can trace better than a white one, since there's no background region to outline, but this depends on whether your tracer uses the alpha channel. If it doesn't, a flat pure-white background is the safer option.

Much of the quality in a colour trace comes from one setting: how many colours the output is allowed to use. Too many and you get a mosaic of near-duplicates with jagged seams. Too few and detail disappears.

As starting points rather than rules: two to four for a logo or icon, six to twelve for a flat illustration. Adjust from there based on what the result looks like at the size you'll use it. For anything genuinely photographic you're past the point where tracing produces a usable file, and you should be thinking about a different approach.

If the generated image is photographic, heavily shaded, or built on gradients, no prompt tweak and no tracer setting will give you a clean vector. The information you want, discrete shapes with hard boundaries, isn't in the image.

For those, trace to get a rough base, then rebuild. Or regenerate with a flatter prompt. Both are faster than cleaning up a file with thousands of paths.

If what you're vectorizing is specifically a logo, this walkthrough goes further into geometry, counters, and when to redraw instead.

Tracer is free for SVG.