FitMyPic

Reference

What actually makes an image file large

Four levers decide how many bytes a picture costs. One of them can cut a file by 90% with no visible change, and it is usually not the one people reach for.

By FitMyPic Editorial Published 2026-10-22 Last reviewed 2026-10-22 How we test

At a glance

  • Pixel count is the strongest lever: halving width and height removes three quarters of the image data.
  • JPEG quality is the weakest above 80: on our photo, q75 to q90 multiplied the file by 2.9 while SSIM rose 0.888 to 0.893.
  • Content matters as much as settings: a flat illustration compressed 4.4x better than a photograph at the same quality, and a screenshot's PSNR never moved (26.7 dB) while its file grew from 170.9 KB to 752 KB.
  • Metadata is real but small: EXIF, thumbnails and colour profiles usually cost tens of kilobytes, not hundreds.

The four levers

Every byte in an image file comes from one of four things: how many pixels there are, how many bits each pixel is allowed to spend, how expensive those pixels are to describe, and what is bolted on beside them. Settings screens usually expose only the second one, which is why the results of "just lower the quality" are so disappointing.

It helps to think in bits per pixel. A picture at 1200 x 900 has 1.08 million pixels. A 150 KB file spends 1.11 bits per pixel; the same image at 1.5 MB spends about 11 bits per pixel. Nothing about the picture changed between those two files, only the budget the encoder was given.

Lever 1: pixel count

This is the lever with the biggest range. Storage scales with the number of pixels, so a 4000 x 3000 photograph carries 12 million pixels where a 1200 x 900 carries one million — a 12x difference in raw data before any compression happens.

That is also why resizing is the first step in almost every size-limit workflow. If a form asks for under 2 MB and you are uploading a 48-megapixel phone photo, no quality slider will get you there gracefully; the encoder is being asked to throw away 95% of the picture's information, and it does that by blurring and blocking. Crop to the aspect ratio you need, then resize to the pixel dimensions the destination actually displays, then spend quality on what is left.

There is one trap: resizing up does not add pixels that were not there. A 600-pixel-wide source stretched to 1200 is a soft 1200-pixel file that costs twice the bytes.

Lever 2: the quality setting, and why it saturates

The quality slider controls how aggressively the encoder quantises the 8 x 8 blocks of frequency data that JPEG and WebP work in. Low settings coarsen the quantisation tables; high settings keep more detail and spend more bits.

The relationship is sharply non-linear, and the curve flattens earlier than most people expect. From our JPEG sweep on a synthetic photo-like test image (1600 x 1200, encoded with libjpeg-turbo):

Quality File size Change vs q75 PSNR SSIM
q40 46.8 KB 0.33x 38.3 dB 0.877
q60 76.3 KB 0.54x 38.6 dB 0.883
q75 141.2 KB 1.00x 38.7 dB 0.888
q80 193.5 KB 1.37x 38.8 dB 0.890
q85 270.2 KB 1.91x 38.9 dB 0.892
q90 414.7 KB 2.94x 38.8 dB 0.893
q95 952.8 KB 6.75x 38.0 dB 0.877

Two things stand out. First, the structural similarity metric moves 0.888 to 0.893 between q75 and q90 while the file grows 2.9x — a difference most people cannot see on a normal screen. Second, q95 is not "better": it is 6.75x the bytes of q75, and both measured metrics get worse, because the encoder starts preserving noise the eye does not reward.

The practical reading: on photographs, quality between 75 and 85 is the band where extra bytes stop buying much. If you need a smaller file than that gives you, change the dimensions instead of continuing to pull the slider down.

Lever 3: content complexity and format

How expensive the pixels are to describe depends on the picture, not the settings. Compress the same encoder at the same quality across different image types and the spread is enormous:

Image type Size at JPEG q85 What it says
Flat illustration 85.4 KB Large areas of identical colour cost almost nothing
Smooth gradient 66.4 KB Predictable variation is cheap
Photo-like texture 270.2 KB Grain and detail resist compression
Screenshot (UI, text, thumbnails) 394.6 KB Text edges are the worst case for JPEG

This is why the screenshot row of the quality sweep is a warning rather than a dial: from q40 to q95 its file grew 170.9 KB to 752 KB — a 4.4x increase — while PSNR stayed at 26.7 dB and SSIM crept from 0.917 to 0.934. Nearly every extra byte went into describing text edges more precisely, and the metric that reflects what people notice (structure) barely responded. Screenshots belong in a lossless or near-lossless format, which is a format decision rather than a quality decision.

Format choice is a second, orthogonal lever. For the same picture, converting from a lossy WebP to PNG re-encodes the data losslessly and typically multiplies the file several times over — our format study measured 7.3x to 18.2x across six image types. Going the other way, from PNG to JPEG at q85, usually shrinks a photograph to a fraction of its size and a flat illustration by less; the benchmark page lists the exact multipliers per type.

Lever 4: metadata and everything bolted on

EXIF blocks, embedded colour profiles, camera thumbnails and XMP blocks ride along inside the file. For a phone photograph this is typically tens of kilobytes — noticeable on a 200 KB web image, irrelevant on a 6 MB one. Stripping metadata also removes the GPS coordinates and device details that many people would rather not publish, so it is often worth doing for reasons that have nothing to do with size.

One category is worth calling out because of scale: embedded previews. Some editors save a full-size preview inside the file, which can double it. If a file is mysteriously much larger than an equivalent export from another program, that is the first thing to check.

The order to work in

  1. Decide the pixel dimensions from the destination, not from the camera. A web page or an app screen has a fixed budget; a print has a physical size and a viewing distance.
  2. Crop to the aspect ratio before scaling, so that resizing does not have to invent or discard proportions.
  3. Resize once, with a quality-oriented filter. Nearest-neighbour downscaling left 6.85 levels of halftone noise where a proper filter left 1.18, and its output even compressed worse afterwards (79.9 KB versus 49.5 KB as JPEG q85).
  4. Set quality in the 75-85 band for photographs, and pick a lossless format for text-heavy images instead of pushing quality up.
  5. Check the result against the actual requirement: dimensions, byte size, and whether the destination accepts the format at all.

What "too large" usually is not

A 4 MB PNG of a flat logo is not "too large" because PNG is inefficient; it is usually an image saved with a 24-bit photographic background it never needed. A 12 MB photograph is not "too large" until you know what it is for: as an archive master it is normal, as a web hero image it is two orders of magnitude over budget.

The measurement for your own image is always available: load it in the resizer at the pixel size you need, watch the output size beside the quality slider, and stop when the file meets the requirement rather than when the slider hits a number someone told you was standard.

Resize, compress and convert images in your browser

Everything happens on your device — no upload, no account.

Open the image tools