Method
Resizing without losing quality: what is actually possible
You cannot resize losslessly. You can keep the loss invisible, and the difference between the two is a short list of habits.
At a glance
- Resizing is always a reconstruction: every output pixel is recomputed, so "lossless resize" does not exist — but visible loss is avoidable.
- Resize once, from the largest original. Every additional resample or re-encode compounds the loss.
- A quality-oriented filter is the single biggest lever: 44-46 dB against a reference versus 31 dB for nearest-neighbour on our 4x downscale test.
- Downscaling removes detail permanently; upscaling cannot restore it, and mixes in halos of its own.
Why "lossless resize" is a phrase to be suspicious of
A resize recomputes a new pixel grid from the old one. Unless the new grid is an exact integer multiple of the old one — and usually not even then — the values in the output are weighted averages of neighbouring input values. Averages do not round-trip: resize 1000 pixels down to 300 and back to 1000, and the result is a blurred approximation, not the original.
So the goal is not zero loss. The goal is to keep the loss below the threshold where it becomes visible, and to avoid the three mistakes that push it over.
Rule 1: start from the biggest, best source you have
Compression and resizing are both one-way. Each pass discards information, and starting from an already-processed file means starting part-way down the slope.
Work from the camera original, the exported PNG, or the largest version you can obtain. Do not start from a chat app's copy, a screenshot of a photo, or a file that has already been sent through a compressor at quality 60. If the only copy you have is a small JPEG, accept that it constrains the result rather than "fixing" it with processing.
Rule 2: resize exactly once
The most common avoidable quality problem is a chain: resize to a large size, resize again to the final size, then let an upload form's tool resize a third time.
Each interpolation step softens a little more. Instead, decide the final dimensions first, then interpolate once, directly from the source. This also means avoiding workflows where an editing app exports an intermediate at a different size than the destination needs.
If you must chain — say the composition is edited after cropping — do the geometric work at full resolution and only then perform the single final resample.
Rule 3: choose the filter deliberately
Measured on a 4x downscale of a noisy test pattern, compared against an area-average reference:
| Filter | PSNR | SSIM | What it leaves behind |
|---|---|---|---|
| Nearest neighbour | 31.66 dB | 0.6028 | Aliasing and moire (6.85 levels of halftone noise) |
| Bilinear | 44.39 dB | 0.9813 | Mild softening, no ringing |
| Bicubic | 46.60 dB | 0.9856 | Slight edge sharpening, mild halos when upscaling |
| Lanczos | 46.52 dB | 0.9835 | Slightly more edge contrast, more halo when upscaling |
For photographs being reduced in size, bicubic and Lanczos are the two defensible answers, and the difference between them is small. Nearest-neighbour should be reserved for pixel art. Bilinear is a fine compromise when a tool offers nothing better, and it has the useful property of never introducing ringing.
Sharpness added after a downscale does not restore detail; it amplifies whatever survived. If the result looks soft, the fix is a smaller reduction — target fewer pixels less aggressively — not a sharpen pass.
Rule 4: do not re-encode what you already encoded
Compression artefacts are not random noise. Re-encoding a JPEG at the same or lower quality re-quantises blocks that already contain artefacts, so blocking and ringing compound instead of averaging out.
Practical consequences:
- Export from the editor directly at the final size and quality. Avoid the "save, then resize, then save again" loop.
- When you must combine resize and compression, do them in one pass in one tool: the resizer and compressor produces a single output rather than feeding one export into another.
- If a destination accepts PNG and the image has text or flat colour, send PNG: it avoids the compounding problem entirely, at the cost of a larger file.
Rule 5: pick dimensions the destination can actually use
A common cause of "my resize looks bad" is resizing to a size that will be resampled again downstream — by a platform's own pipeline, by a template, or by a print shop. Reaching the destination's native size, or an exact multiple of it, gives its software nothing to do.
For screens, that means the CSS pixel size multiplied by the display density you intend to support: 1080 wide for a typical feed image, 1500-2000 wide for a full-width hero on a high-density laptop, and no more than that unless the image is a zoomable detail. For print, multiply the physical size in inches by the pixels per inch you actually need — 300 is the conventional figure for close viewing, and 150 is often indistinguishable at arm's length.
Rule 6: verify at 100%, then stop touching it
Open the output at 100% zoom and look at the parts of the image with the most high-frequency detail: fabric, foliage, hair, small text. If those areas look acceptable at the size they will be viewed, the job is done.
Do not keep iterating. Every additional export adds a generation of loss, and successive small improvements in one area usually cost more in another. Keep the source file, note the settings that worked, and stop.
What this looks like as a checklist
- Locate the largest original.
- Decide final pixel dimensions from the destination, not from the file.
- Crop to the target aspect ratio at full resolution.
- Resample once, with bicubic or Lanczos.
- Compress in the same pass if a size limit applies, at quality 75-85 for photographs.
- Inspect at 100% on the busiest part of the image.
- Save, and keep the original untouched.
Resize, compress and convert images in your browser
Everything happens on your device — no upload, no account.