Image Upscaling, Explained

After reading this you will know how enlarging an image actually works, why Lanczos resampling suits photos and Scale2x suits pixel art, and how to judge whether an upscale looks good or just looks big.

What upscaling really does

An image is a grid of pixels. A 400x300 photo holds 120,000 color samples. When you upscale it 2x to 800x600, the new grid has 480,000 pixels, four times as many. The camera never recorded those extra 360,000 samples, so the computer has to invent them from the ones it has.

That inventing step is called resampling. Every upscale answers one question repeatedly: given the known pixels around this new empty position, what color belongs here? The differences between upscalers are entirely in how they answer that question.

Here is the hook. Take a single pixel at position (10, 10) in the small image. After a 2x enlargement it does not simply become a pixel at (20, 20). Instead the new grid asks for colors at (20, 20), (21, 20), (20, 21) and so on, and each of those falls between the original samples. The upscaler estimates each one. A good estimate keeps edges sharp. A poor one smears them.

When to use each method

The right algorithm depends on the source, not on the size you want.

Lanczos resampling

Best for photographs, scans and any image with smooth gradients. It blends nearby pixels with a weighting curve, so it reconstructs soft transitions without hard blockiness.

Scale2x and Scale3x

Best for pixel art, sprites, low-resolution game graphics and simple line icons. These algorithms look at the shape of each pixel's neighbors and keep straight edges straight instead of blurring them.

Use Lanczos on a family snapshot. A 2x enlargement of a face stays natural because skin, hair and sky are all smooth. Use Scale2x on a 32x32 game sprite, because Lanczos would turn its crisp black outline into a gray haze.

The math behind Lanczos

Lanczos resampling weights each nearby source pixel by a windowed sinc function. For a scaling factor, each output pixel maps back to a fractional position in the source, and pixels within a radius a contribute according to this kernel.

Here x is the distance from the output sample to a source pixel, measured in source pixels. The term \text{sinc}(x) = \sin(\pi x)/(\pi x) is the ideal interpolation shape, and \text{sinc}(x/a) is the window that limits how far the influence spreads. The radius a is usually 2 or 3. With a = 3, six source pixels along each axis feed every output pixel.

The key feature is that L(x) goes negative for some distances. A pixel one unit past the center gets a small negative weight. Those negative lobes are what sharpen edges: a bright pixel next to a dark one pulls its neighbor slightly darker, which keeps the boundary crisp instead of washed out. Bilinear interpolation, by contrast, uses only positive weights and always blurs.

How Scale2x decides

Scale2x turns each source pixel into a 2x2 block of four output pixels. It does not average anything. Instead it compares the four orthogonal neighbors (up, down, left, right) and chooses each of the four output pixels based on whether an edge runs diagonally through the corner.

Call the center pixel C, with neighbors A (up), B (right), D (left), F (down). The top-left output pixel becomes D if D == A and D != F and A != B, otherwise it stays C. The other three corners follow the same rule rotated. The effect is that a diagonal line in the source becomes a smoother staircase without any new colors appearing. Every output pixel is an exact copy of some input pixel, so the palette never changes and the art never blurs.

A worked example

Comparing methods on the same image

The table shows how a 200x200 source behaves under different factors and methods. Pixel counts are exact; sharpness is a qualitative note.

Notice that Lanczos 2x and Scale2x 2x produce the same 160,000 pixels. The size is identical; only the pixel values differ. On a photo Scale2x would leave visible steps, and on a sprite Lanczos would leave visible blur.

Common mistakes

The first mistake is expecting sharp detail from a blurry source. Upscaling a 640x480 webcam frame to 2560x1920 gives you 4x the pixels and the same softness. The blur simply covers more screen area.

The second mistake is using Lanczos on pixel art. A 16x16 icon upscaled 4x with Lanczos has 4096 output pixels, but its clean edges dissolve into anti-aliased fuzz. Scale2x twice, or Scale2x then Scale2x again, keeps every edge sharp.

The third mistake is stacking upscales to chase quality. Going 2x then 2x is not the same as 4x once. Each pass reinterpolates already-interpolated pixels, so errors compound. Do the full enlargement in one step when you can.

The fourth mistake is ignoring file size. Four times the pixels usually means a much larger file. After upscaling, run the result through the Image Compressor to trade a little quality for a smaller download.

Related tools

Upscaling is one step in a wider workflow. To change dimensions without enlarging, or to fit an exact target, use the Image Resizer. To trim before or after, the Aspect-Ratio Crop locks to presets like 1:1 and 16:9. If your upscaled art is destined for a browser tab, the Favicon Generator produces every required size from a single source. Your images stay in the browser and are never uploaded.

Related tools

Image Compressor

Shrink JPEG, WebP and PNG files with a live size preview.

Image Resizer

Resize one image or a whole batch to exact dimensions.

Aspect-Ratio Crop

Crop images freely or locked to presets like 1:1 and 16:9.

Favicon Generator

Every favicon size plus manifest from one image.

Seam Carving (Content-Aware Resize)

Shrink or stretch a photo without squashing it: remove the boring pixels, keep the subject.

EXIF Viewer & Remover

See and strip photo metadata: camera, GPS, dates.