Upscale First or Vectorize First? The Order Changes Everything

AI upscalers promise to recover detail from low-resolution images, and they can be impressive on photographs. But vectorization is not photography. When a vectorizer traces an upscaled image, it is tracing the upscaler's best guess at what the edges should look like — and that guess can include hallucinated detail that was never in the original. The result: wavy paths where the true edge was straight, extra nodes where the original had none, and corners that shift by several pixels from their real location.

This guide explains when upscaling before vectorizing helps, when it makes things worse, and how to tell the difference.

What AI upscalers actually do

Modern AI upscalers are trained on millions of high-resolution images and learn to predict what detail should exist at scales the source image does not contain. When you feed in a 100×100 icon and ask for a 400×400 output, the upscaler does not simply interpolate between pixels — it guesses what edges, textures, and fine details would be present if the image had been captured at higher resolution in the first place.

For photographs, this is often a good trade. The upscaler adds plausible detail that makes the image look sharper to the human eye. For logos, line art, and anything with mathematically straight edges, it is a problem. The upscaler does not know that the edge is supposed to be perfectly straight; it only knows what natural images usually look like. So it adds tiny wobbles, sub-pixel shifts, and smoothing that a human would not notice but a vectorizer will trace as real geometry.

When upscaling helps: uniformly soft sources

Upscaling before vectorizing makes sense when the source image is uniformly soft — low resolution, out of focus, or heavily compressed — and there is no higher-quality version available. In this case, the upscaler can sharpen edges and make the boundary between foreground and background clearer, which gives the vectorizer a better target to trace.

Old scans and low-resolution downloads

A 150 PPI scan of a printed logo from the 1980s. A 72 DPI screenshot of a web graphic. A thumbnail downloaded from a stock image site. These sources all have blocky, stair-stepped edges because the original resolution was too low. An AI upscaler can smooth those stair-steps into curves, which the vectorizer can then fit more cleanly.

Soft focus and motion blur

If the source image has edges that blur across three or four pixels, an AI upscaler trained on sharp images will try to sharpen them. This can help the vectorizer decide where the boundary lands, especially if the blur is uniform across the whole image. The trade-off: the upscaler might sharpen some edges correctly and hallucinate others incorrectly, so you need to inspect the traced output carefully.

When upscaling hurts: compression artifacts and clean edges

Upscaling makes things worse when the source image has JPEG compression artifacts, anti-aliasing halos, or already-clean edges that do not need help. In these cases, the upscaler amplifies the artifacts or introduces new ones, and the vectorizer traces the amplified noise as if it were real detail.

JPEG compression artifacts

JPEG compression creates 8×8 pixel blocks with ringing artifacts near edges. An AI upscaler sees those artifacts and tries to make sense of them, often by turning them into textured regions or wavy lines. The vectorizer then traces those hallucinated features. The result: a logo that was a simple four-path shape becomes a 40-path mess. For JPEG sources, clean up the artifacts before tracing, do not upscale.

Already-clean edges at moderate resolution

A 300 PPI scan of a printed logo has clean edges. Upscaling it to 1200 PPI does not add real detail; it adds the upscaler's interpretation of what a 1200 PPI version might look like. If the original edges were straight, the upscaled version might have micro-wobbles that the vectorizer will trace. The rule: if the source already has clean edges at 250–400 PPI, trace it as-is.

How to test whether upscaling helped

The best way to know if upscaling improved the traced result is to trace both the original and the upscaled version, then compare the outputs. Open both traced SVGs in a vector editor at the same scale and overlay them. If the upscaled version has cleaner paths with fewer nodes and straighter edges, the upscaling helped. If it has more nodes, wavy paths, or shifted corners, it hurt.

Node count as a proxy for quality

A simpler test: count the number of nodes in each traced output. For logos and line art, fewer nodes usually means cleaner tracing. If the upscaled version has 50% more nodes than the original, the upscaler probably hallucinated detail. If it has the same number of nodes or fewer, it might have helped.

Spot-check straight edges

Zoom in on a section of the traced output that should be a perfectly straight line. If the line is made of three or four Bezier segments with control points that deviate from a straight path, the source had noise or the upscaler added wobbles. If the line is a single segment or two segments that form a clean straight edge, the source was good.

Upscaling settings that matter for vectorization

Most AI upscalers offer settings for sharpness, noise reduction, and detail enhancement. For vectorization, you want moderate sharpness and aggressive noise reduction. High sharpness settings can create overshoot halos that the vectorizer will trace as separate shapes. Low noise reduction leaves specks that become extra paths.

Recommended settings

Use the photo or standard preset if available, not the illustration or artwork preset. The illustration preset is trained to preserve hand-drawn lines and textures, which is the opposite of what you want. Set noise reduction to medium or high. Set sharpness to low or medium. Avoid any setting labeled texture enhancement or detail recovery, because those add synthetic detail the vectorizer will trace.

Alternatives to upscaling: preprocessing the low-res source

Before reaching for an upscaler, try preprocessing the low-resolution source to make its edges cleaner. These techniques do not add resolution, but they can make the existing resolution more useful.

Bilateral filter

A bilateral filter smooths regions of similar color while preserving edges. This removes noise and compression artifacts without blurring the boundaries the vectorizer cares about. Most image editors have this filter under Noise Reduction or Surface Blur. Apply it lightly before tracing a noisy source.

Unsharp mask

An unsharp mask accentuates edges by boosting contrast at boundaries. For a soft-focus source, a light unsharp mask can make edges clearer without introducing the hallucinations an AI upscaler might add. Keep the radius small (1–2 pixels) and the amount moderate (50–100%).

Levels adjustment

Stretch the histogram so the brightest pixels are white and the darkest pixels are black. This maximizes contrast and makes color quantization simpler during tracing. For a scanned document with gray instead of white background, this step alone can eliminate a whole layer of spurious paths.

Special case: pixel art and low-bit graphics

Pixel art is intentionally low resolution. Each pixel is part of the design. Upscaling pixel art before vectorizing defeats the purpose, because the upscaler smooths the hard pixel boundaries into curves. For pixel art, either trace it at the original resolution and accept the blocky result, or use a nearest-neighbor upscale (not AI) to make the pixels bigger without adding anti-aliasing.

Tools and workflows

Popular AI upscalers include Topaz Gigapixel, waifu2x, Real-ESRGAN, and the upscaling features built into Photoshop and GIMP. For vectorization, avoid upscalers trained specifically on anime or illustrations, because they add stylistic artifacts. Use general-purpose photo upscalers with noise reduction enabled.

For batch processing, command-line tools like Real-ESRGAN or waifu2x-ncnn can upscale folders of images before handing them to a PNG-to-SVG converter. Test the pipeline on a few sample images first to confirm that upscaling improves the traced output before committing to a large batch.

Related resources

For guidance on choosing the right source resolution before tracing, see Best Resolution for Vectorizing. For the tracing pipeline itself, see Image Tracing Explained. For optimizing the traced output, see SVG Optimization Guide. For preparing sources with compression artifacts, see File Formats Guide.