Resizing looks like the simplest possible image operation: you have 4000 pixels across and you want 1200. In practice it is where a surprising amount of image quality is lost, because the arithmetic of turning one grid of pixels into a differently-sized grid has good answers and bad ones, and the bad ones are faster.
The symptom is familiar. A photo that looked crisp at full size comes back from a resizer looking slightly out of focus, or with fine detail turned into a shimmer of stair-stepped edges. Nothing was cropped and nothing was compressed harder — the resampling itself did that.
Why the filter is the whole game
When you shrink an image, each output pixel has to represent an area of several input pixels. The cheapest method picks the nearest input pixel and discards the rest, which is fast and throws away most of the information — that is where jagged edges and shimmering detail come from. A slightly better method averages the four neighbours, which is smoother but soft, because averaging is a blur.
The methods professional tools use — Lanczos and its relatives — weight a wider neighbourhood with a function that deliberately overshoots slightly at edges. That overshoot is what keeps a resized edge looking like an edge instead of a gradient. It costs more computation per pixel, which is why cheap resizers do not use it, and it is the single largest factor in whether a downscaled photo looks sharp.
Filum's Resize Image uses a Lanczos3 filter for downscaling, so edges stay clean rather than going soft or jagged. For JPG and WebP output the re-encode quality is then tuned per image against a measured perceptual score, which is shown with the result — so the resize and the re-compression are both accounted for, rather than the second one quietly undoing the first.
Exact pixels, percentages, and the aspect ratio trap
Two different jobs hide behind one button. Sometimes you need an exact size because something demands it — a profile picture that must be 400 by 400, a form that rejects anything over 1024 pixels wide, a print layout with a fixed frame. Sometimes you just need the file to be smaller and any reasonable reduction will do, which is what a percentage is for.
The trap is the aspect ratio. If you type both a width and a height that do not match the original proportions, something has to give: either the image is stretched, or it is cropped, or it is padded. Tools differ in which one they pick and most do not tell you. Stretched faces and squashed logos almost always come from a resize where both boxes were filled in without thinking.
The safe default is to lock the ratio and enter one dimension, letting the other follow. Filum locks it by default and lets you unlock it deliberately when you genuinely want to force both, with a preview showing exactly what you will get before anything is written.
Making an image larger
Enlarging is a fundamentally different problem from shrinking, and it is worth being clear-eyed about it. When you shrink, you are choosing which of the available detail to keep. When you enlarge, there is no additional detail available — the extra pixels have to be invented from their neighbours. A good filter makes the invention smooth and plausible; it cannot make it real.
This matters because forms sometimes demand a minimum dimension that your file does not meet, and the honest solution is to enlarge and accept a slightly softer picture rather than to expect magic. Filum allows up to 4× per side and labels the result as enlarged rather than describing it as enhancement.
If what you actually want is a genuinely better large version — a small photo or illustration made bigger and cleaner rather than merely bigger — that is a different technique, using a model trained on what detail typically looks like at higher resolution. Filum's Upscale Image does that job, also on your own device.
Formats, transparency, and where the work happens
PNG output stays lossless: the pixels are resampled once and then stored without any further quality loss, with the alpha channel intact. That makes PNG the right target for logos, screenshots, and anything with transparency. JPG and WebP are re-encoded, which is why the measured quality of the delivered file is worth showing rather than assuming.
Filum's Resize Image supports JPG, PNG, and WebP, and tells you immediately if you drop a format it does not handle rather than failing halfway through. The resampling runs inside your browser on your own device, so the picture is never transmitted to a server to be resized — which matters more than it sounds for the images people actually resize, which are frequently identity documents, screenshots, and photographs of other people.