FitMyPic

Workflow

How to resize an image: settings, aspect ratios and resampling

Resizing looks like a single number to type. It is really four decisions, and the last one — the filter — is invisible in most tools and visible in every result.

By FitMyPic Editorial Published 2026-10-22 Last reviewed 2026-10-22 How we test

At a glance

  • Resizing changes the pixel grid; the filter decides how each new pixel is calculated from its neighbours.
  • Downscaling 4x, a proper filter left 1.18 levels of halftone noise where nearest-neighbour left 6.85 — visible moire in the second case.
  • Bicubic scored highest against an area-average reference (46.6 dB, SSIM 0.9856) on our test pattern; bilinear was close and roughly half the work.
  • Enlarging 4x never adds detail, and sharp kernels add halos instead: bicubic overshot a hard edge by 10/255, Lanczos by 15/255.

What a resize actually does

Every resize is a reconstruction: the output grid rarely lines up with the input grid, so each output pixel is computed from a weighted neighbourhood of input pixels. The weights come from the resampling filter — the kernel — and that choice, not the number you typed, determines how the result looks.

That is why two tools can both produce "1200 x 900" and give visibly different files.

Decision 1: the pixel dimensions

Set dimensions from the destination. A platform that displays at 1080 wide does not need 4000 pixels of width; a printed page needs roughly 300 pixels per inch of paper, so an A4 print at full quality wants about 2480 x 3508 pixels. Everything between those two cases is a judgement, and the tool on this site reports the resulting pixel count so you can see it while you type.

Percentages are convenient but hide what is happening: "50%" of a 4000-pixel image is still 2000 pixels, which is often far more than the destination will ever show.

Decision 2: aspect ratio, and what to do when it does not match

Sources and destinations rarely share a ratio. Three strategies exist, and they are not interchangeable:

Strategy What it does Cost
Fit (letterbox) The whole image is kept, blank space fills the gap Added borders; the subject is smaller
Crop to fill Fills the frame, edges are cut off You lose part of the picture
Stretch Forces the image into the frame Distortion — almost never acceptable

Fit and crop both preserve geometry; stretch does not, which is why "why does my picture look squashed" almost always traces back to a stretch. Crop deliberately: preview where the cut lands, especially with faces, text or logos near the edge. For profile pictures and thumbnails that are cropped again by the platform, keeping the subject centred with margin is the safer habit.

Decision 3: up or down, and by how much

Downscaling removes information; upscaling invents positions for pixels that never existed. Both are legitimate, but they behave differently.

Downscaling is where filters matter most. Our test pattern downscaled 4x and compared against an area-average reference gave:

Method PSNR vs reference SSIM Halftone noise
Nearest neighbour 31.66 dB 0.6028 6.85
Bilinear 44.39 dB 0.9813 1.18
Bicubic 46.60 dB 0.9856 1.44
Lanczos 46.52 dB 0.9835 1.57

Nearest neighbour is not a resampling filter in any useful sense for photographs: it samples one source pixel per output pixel and throws the rest away, which is why its output kept 6.85 levels of halftone noise where bilinear kept 1.18. It is still the correct choice for one job only — scaling pixel art or an image where every pixel must remain an exact original value.

Upscaling is bounded by arithmetic. At 4x on a hard edge, bicubic overshoots the edge by 10 levels of 255 and Lanczos by 15, producing the bright and dark halo photographers know as ringing. Bilinear and nearest produce none, at the cost of a softer result. If you are enlarging for print and the destination is not enormous, the honest options are a modest enlargement (up to roughly 2x for on-screen results) or a different source file.

Decision 4: the filter your tool is actually using

This is the setting most tools hide. Where it is exposed, the rule of thumb that matches our measurements is simple: use a high-order filter (bicubic or Lanczos) for photographs, use nearest-neighbour only for pixel art, and prefer the tool's "high quality" path when the tool wraps canvas resampling, because the browser default is tuned for speed.

Speed is rarely the constraint for still images. In our runs Lanczos took about 2.4x the time of bilinear for a 4-megapixel source — a difference measured in a fraction of a second, on a decision whose visual result persists.

After the resize

Resizing is normally the first of two or three steps, not the last. Once the pixel dimensions are right, compress if a size limit applies, and check the result at 100% zoom on the edges: text, hair and straight high-contrast lines. Then save once, from the resized canvas — never re-open a compressed export to resize it again.

If you resized down and the result looks soft at full size, that is expected: the detail was removed by the downscale. Reducing the dimensions further will not bring it back, and enlarging will not either.

Resize, compress and convert images in your browser

Everything happens on your device — no upload, no account.

Open the image tools