Image scaling algorithms, compared on your own picture

Pick a crop, pick a scale, and the same pixels are rendered by four methods at once. Hover to compare the same spot everywhere; download the full picture with the one you prefer.

Click or drag on the picture to choose the crop.

Scale
Magnify panes
Source—
Scaled0 × 0 px
Crop shown0 × 0 at (0, 0)

Use your own picture

or paste from the clipboard · JPEG, PNG, WebP, GIF, BMP, AVIF

…
…
…
…

Each pane is the same 0 × 0 px crop of the picture scaled to 400%, computed by the real engine and shown at 1× with no smoothing. The bar under each pane compares edge contrast between panes: higher is crisper, though very high on an enlargement can mean ringing.

The methods

All are separable filters applied in linear light unless you choose the sRGB variant:

  • Nearest neighbour — copies the closest pixel. Blocky up, aliased down; perfect for pixel art.
  • Box / area — averages exactly what each output pixel covers. The fairest reduction; blocky when enlarging.
  • Bilinear — a triangle filter over two pixels each way. Smooth, soft.
  • Bicubic Catmull-Rom — the sharp cubic. Slight halos on hard edges.
  • Bicubic Mitchell — a softer cubic designed to balance blur, ringing and blockiness.
  • Lanczos-2 / Lanczos-3 — windowed sinc with two or three lobes. Sharpest; can ring beside hard edges.

What to look for

On enlargements, compare edges: blocky (nearest), soft (bilinear), sharp with a halo (Catmull-Rom, Lanczos). On reductions, look at the zone plate and gratings of the test chart: false rings and patterns are aliasing. Compare “Lanczos-3” with “Lanczos-3 · sRGB” on the checkerboard patch to see why gamma-correct scaling matters.

Questions

Which scaling algorithm is best?

There is no single best: Lanczos-3 is the sharpest general choice for photos, Mitchell is the safest for big enlargements, Box is the fairest for reductions of graphics, and Nearest is the only correct choice for pixel art. That is why the lab exists — the right answer depends on the picture.

Is bicubic the same as Catmull-Rom?

“Bicubic” is a family. Catmull-Rom (B=0, C=0.5) is the sharp member most software means; Mitchell (B=C=1/3) is softer with almost no ringing. Both are here.

Are the panes accurate, or previews?

Accurate. Each pane is computed by the same engine at the same scale as a full-size job, for just the selected region.

What does “linear light” mean, and why is it on by default?

Pixel values in a JPEG or PNG are gamma-encoded: value 128 is not half as bright as 255, it is about a fifth. Averaging those numbers directly — what most resizers do — darkens fine bright detail and muddies colour edges. With linear light on, every pixel is converted to physical light before it is weighted and converted back afterwards. The test chart’s one-pixel checkerboard shows the difference: reduced in linear light it matches the light grey patch beside it; reduced the naive way it matches the dark one.