Why smaller is harder than it looks
Making a picture smaller throws pixels away, and which ones survive decides whether a striped shirt stays striped. A kernel evaluated at the output’s rate cannot see detail finer than two output pixels, so below half size that detail does not vanish — it folds into a coarser false pattern. Averaging whole blocks of the source first removes it before it can fold, and the kernel finishes the job on what remains.
| Factor | Pixels | Area-average first | At 300 dpi |
|---|---|---|---|
| ×0.75 (75%) | 4500 × 3000 | no | 38.10 × 25.40 cm |
| ×0.50 (50%) | 3000 × 2000 | yes | 25.40 × 16.93 cm |
| ×0.25 (25%) | 1500 × 1000 | yes | 12.70 × 8.47 cm |
| ×0.10 (10%) | 600 × 400 | yes | 5.08 × 3.39 cm |
Limits worth knowing
- Every reduction discards recorded detail. Keep the original; going back up from a reduced copy cannot recover it.
- Two reductions in a row round and filter twice. Go straight to the final size from the original.
- In resample mode the browser writes a new container, so the camera's Exif block and any colour profile do not survive. The resolution tag is written again for JPEG and PNG.
Questions about downsizing
What is moiré, and why does shrinking a picture cause it?
Moiré is a false pattern — waves or bands — that appears when fine repeating detail such as fabric, brickwork or a screen door is sampled too sparsely. A reduction samples the picture at a lower rate, and below half size an ordinary kernel samples too sparsely to see the fine pattern correctly, so it folds into a coarser one that was never there.
What does this page do differently at 50% and under?
It averages whole blocks of source pixels first — an integer area-average pass — and only then runs the chosen kernel for the remainder. The averaging removes exactly the detail that would fold, so what is left can be resampled cleanly.
Which kernel should I use going down?
Lanczos keeps the most sharpness and is the automatic choice for reductions. On line art and interface screenshots it can leave a thin bright rim beside hard edges; bicubic gives up a little acuity and does not ring. The kernel strip under the stage shows every kernel on your own picture.
Will downsizing reduce the file size in KB?
Almost always, because there are fewer pixels to store, but the exact length depends on the format and quality you write. The written length is read back from the saved file and shown with its SHA-256. To hit a limit in kilobytes, the size-limit page searches quality first and pixels second, and says which moved.
Should I sharpen after downsizing?
Often a little, yes. A reduction averages away some edge contrast that was genuinely in the picture, and the downscale sharpening control restores it with a small unsharp mask after the resample. Leave it at zero if the result is going to be scaled again by a page.
Other ways to set the size
Given a pixel box instead of a ratio? Set it directly, and choose what happens when the shapes differ.
- Change image size in pixelsA box in pixels: fit inside, fill and crop, or stretch, with the factor printed.
- Scale by percentA ratio in, a pixel count out.
- Image resamplingA percentage-led pixel resize with the interpolation cost stated before export.
- Choosing a resampling kernelNearest, bilinear, bicubic and Lanczos compared, with linear light and alpha.



