Your All-in-One
Online Toolbox

Upscaling means inventing pixels that were never captured, and the algorithm decides how those pixels are estimated from their neighbours. All three methods here are interpolation filters that compute each output pixel as a weighted average of nearby source pixels, differing in how many neighbours they consider and how the weights are shaped.

Bilinear takes the four nearest source pixels and blends them linearly by distance. It is the cheapest to compute and produces smooth but noticeably soft results, because the linear weighting blurs edges.

Bicubic uses a four-by-four neighbourhood, sixteen source pixels, with the Keys cubic weighting function. The cubic curve holds edges better than a straight line, giving a sharper result than bilinear at moderately higher cost.

Lanczos-3 uses a windowed sinc function with a radius of three, drawing on a six-by-six neighbourhood of thirty-six source pixels. Sinc is the theoretically ideal reconstruction filter, and windowing makes it practical. It preserves fine detail and edge definition best, at the highest cost, which is why it is the default and also the slowest.

Resampling is applied separately in the horizontal and vertical directions, and the result is drawn to a canvas and encoded in your chosen format.

Take a 640 by 480 screenshot and select the 3x scale factor. The output dimensions preview shows 1920 by 1440, since each dimension is multiplied by three.

The pixel count is where the cost becomes clear. The source holds 307,200 pixels. The output holds 2,764,800, which is nine times as many, because area scales with the square of the linear factor. Every one of those 2.76 million output pixels must be computed, and under Lanczos-3 each one is a weighted sum over a six-by-six neighbourhood, so roughly 36 multiply-and-add operations per channel per pixel. That is why the reported processing time rises steeply with the scale factor, and why 4x on a large photograph takes appreciably longer than 2x on a small one.

Only about 11 percent of the output pixels correspond to an original sample; the rest are interpolated estimates. This is the fundamental limit: the result holds no more real detail than the 640 by 480 source did. A good filter buys a cleaner, less blocky presentation of existing detail, not new information.

Saved as PNG at 1920 by 1440, the file is several times larger than the original despite carrying no extra detail.

Set expectations correctly. Interpolation cannot reconstruct detail that was never recorded. If a face is eight pixels wide in the source, no resampling filter will produce recognisable features at 4x. Text that is illegible in the original stays illegible, just larger and smoother. This tool is not an AI upscaler and makes no attempt to hallucinate plausible detail; it applies a deterministic mathematical filter, which is why it gives the same result every time.

Where it works well is enlarging already-clean sources: line art, logos, diagrams, screenshots at moderate factors, and reasonably sharp photographs that simply need to fill a larger frame. Use the compare view at full size rather than judging from a thumbnail, since interpolation differences are invisible when the image is scaled down for display.

Existing flaws are magnified too. JPEG artefacts, sensor noise and moiré all become larger and more obvious, so start from the cleanest source available and prefer an untouched original.

On output format, choose PNG for graphics, screenshots and anything with flat colour or transparency, since it is lossless. For photographs, JPEG or WebP at 85 to 95 percent avoids adding fresh compression damage on top of the enlargement.