Image Tools

For photographs, AVIF is roughly 40–50% smaller than JPEG at matched visual quality and WebP lands around 25–30% smaller, with AVIF taking noticeably longer to encode. JPEG remains the safe default when a file has to open in anything, including a decade-old print workflow. PNG is not a photo format — it is lossless, so a photograph saved as PNG is often five times larger than the same image as JPEG at quality 85, with no visible benefit.

PNG earns its place with flat colour, sharp edges and transparency: screenshots, logos, diagrams, UI assets. If the artwork is vector to begin with, SVG beats all of them, because it stays sharp at any size and usually weighs a few kilobytes. HEIC is what an iPhone produces by default — efficient, but poorly supported outside Apple’s ecosystem, which is why converting to JPEG is usually the first step when a photo needs to go anywhere else.

JPEG quality is not a percentage of anything. The number selects a set of quantisation tables that decide how aggressively frequency coefficients get rounded away, and the perceptual curve has a knee: between 95 and 85 the file shrinks a lot and almost nothing changes visually, while below about 70 blocking and ringing become obvious around edges and in flat gradients. For most photography, 75–85 is the useful range.

Chroma subsampling is the setting nobody mentions. At 4:2:0 the encoder keeps full brightness detail but stores colour at half resolution in each direction, which is invisible in a landscape and very visible on red text or a saturated logo edge. Use 4:4:4 when an image contains coloured type or fine graphic detail. And remember that re-saving a JPEG re-quantises it every time, so repeated edit-and-save cycles degrade the file even at quality 95.

Downscaling well means resampling with enough neighbouring pixels to avoid aliasing. Browser canvas scaling is fast but simple, so reducing a 4000-pixel image to 400 pixels in one step can leave stair-stepped edges on fine detail; halving repeatedly, or using a proper Lanczos pass, gives cleaner results. Export at twice your display dimensions if the image needs to be sharp on a high-density screen, then compress harder to keep the file reasonable — a 2x image at quality 70 usually beats a 1x image at quality 95 in both size and appearance.

AI upscaling is a different operation entirely. A conventional resize interpolates between the pixels you have; a model hallucinates plausible detail that was never captured. That is excellent for making an old photo look presentable and completely unsuitable for anything evidential — text gets rewritten into confident nonsense, and faces drift toward whatever the model considers a typical face. Never upscale a document, a licence plate or a medical image and treat the result as real.

A photo straight off a phone carries EXIF: GPS coordinates accurate to a few metres, the exact capture timestamp, camera model and serial number, and often a thumbnail generated before you cropped the image. That last one has exposed things people thought they had cropped out. Most large social platforms strip EXIF on upload, but forums, messaging apps that send files as documents, cloud drives and your own website generally do not.

Strip metadata before publishing anything that came from a camera, and check what survives conversion — converting HEIC to JPEG carries the EXIF block forward intact unless you explicitly remove it. Dominant-colour extraction is the harmless counterpart: it reads pixels and returns a small palette you can use for a placeholder background, a chart accent or a matching theme, without touching the file at all.