Optimize images for the web

Compress JPG does one thing: re-encode at the smallest size that still looks the same. This page also sharpens and resizes, and it does all three inside one decode. That matters more than it sounds. Running three separate tools means the image is encoded and decoded again at every handover, and each lossy round adds noise to the exact picture you are trying to keep clean. Here the image is decoded once at the start and encoded once at the end; everything in between happens on the same uncompressed pixels in memory.

Because resizing softens. Every time an image is scaled down — by us, by your camera, by WordPress generating thumbnails — neighbouring pixels get averaged together, and fine detail turns slightly mushy. Sharpening puts back the definition that averaging took out. It runs before the resize on purpose: sharpening a large image and then shrinking it lets the resize smooth away any overshoot, while sharpening after the resize amplifies the very artefacts the resize just created.

No, and it is worth knowing why. Sharpening adds fine detail, and fine detail is exactly what costs bytes in a compressed image. On an already-compressed photo we measured Light at about +3%, Medium at +8% and Strong at +14%. We keep the sharpened version anyway, because you asked to change the picture — handing back the original would be ignoring the request, not protecting you. With sharpening off, this tool never returns a file larger than the one you uploaded.

Medium for most photographs. On our own test image Light raised measured edge energy by 5% against 4% for no sharpening at all — a difference small enough to be lost in re-encoding, which is to say you will probably not see it. Medium raised it by 39%, which is clearly visible on eyelashes, eyebrows and hair without looking processed. Strong more than doubles it and starts showing halos around high-contrast edges and grain across skin; keep it for soft scans and screenshots, not faces.

It can make a soft image look crisper, but it cannot invent detail that was never captured. Sharpening works by increasing contrast at edges that already exist. If the detail is gone — out of focus, motion blurred, or destroyed by heavy compression — no amount of sharpening brings it back, and pushing harder just makes the blur look harsh. For genuinely small images there is Upscale Image, which uses a model to add plausible detail, with all the caveats that implies.

It is SSIM — a structural comparison between the result and your original, measured at full resolution rather than eyeballed. 1.0000 means pixel-identical. Around 0.99 is where almost nobody can tell the two apart. Below 0.98 people start noticing. It is the honest other half of the savings figure: crushing an image gives a wonderful percentage and a poor score, and you should see both numbers before deciding.

PNG is lossless, so a photograph saved as PNG has very little redundancy left for us to remove. The only way to shrink it much is to cut down to 256 colours, which puts visible banding across skies and skin — we measure that and refuse when it would. The real fix is a different format: switch the extra format to WebP and you will typically see the file drop by around 90% with no visible change.

Files up to 50 MB are accepted, and a batch is handled as separate jobs that are zipped together at the end. Every upload is stored under a random name, never served from a public folder, and deleted 24 hours after the job starts — the finished file goes at the same moment. We never share them and there is no account they could be tied to. The one thing we do is compare the original against the result in our own admin area, to be sure the compression has not softened the picture.