How it measures

ImgTune looks at three levers for each image: resolution (the pixels the browser actually needs), quality (the encoder setting) and format (for example AVIF or WebP instead of JPEG). Measurements are per page view, per viewport, and quality is scored with SSIMULACRA2.

Viewports and page view

Each page is loaded at six viewports and scrolled to the end. Each image is recorded with the size it renders at, the device pixel ratio, and the pixels it needs (rendered width x dpr, capped at the largest file the site offers).

The estimated saving

The estimated saving is the one number to act on. It is the share of a page view's image bytes that a better file would save. It has two parts:

• A smaller file the site already has. The site's own smaller srcset candidates, each fetched and measured, so this part is the real byte difference.
• A re-encode. The image encoded locally at a lower quality or in a newer format (or resized, for a plain <img>). A re-encode counts only when SSIMULACRA2 says it looks the same as the original (see below). These are estimates, and they are labelled as such.

The CLI prints the estimate on its saving exact + est. line. Its saving exact line is the first part alone, and the difference is the re-encode part. A figure the CLI marks with ≈ is approximate: the file is above the size limit for a full match.

A page's saving is the share of all its image bytes per page view. It is the mean over the viewports of each viewport's share, and a site's figure is the mean over its pages. The same method is used for the production study.

Quality: SSIMULACRA2

Each candidate is judged with SSIMULACRA2 against the original. The score runs from 100 (identical) downward.

The default setting is relaxed: a candidate passes when its score is at least 70, and no more than 2 points below the current quality's score. strict uses 80 and 1 point. The CLI reports a delta 4 column with --aggressive, beside the main numbers. Quality is judged at the widths the image is shown at.

ImgTune only accepts a saving when SSIMULACRA2 says nobody can see the difference.

Sharpness

An image smaller than the slot it fills is undersized. It is reported as sharpness growth, in KB per view, apart from the savings. Growth from undersized images is never subtracted from other images' savings, and an undersized image contributes no saving itself.

Fix hints

Every image with a saving, or with a sizes value that differs from the current one, gets a fix hint. The hint gives the sizes, quality and format to use, and two byte numbers: the bytes the fix saves per page view, and the bytes it adds for sharpness. The two are never netted. A scan's hints stand on the measured-page rung. See Fix it with your agent for the other rungs.

The LCP image

The largest contentful paint is the image the first paint waits for. ImgTune reads it at each viewport before the page is scrolled, so a lazy image that loads on scroll cannot take its place. The LCP image is the first line of every page's report and the first hint of the fix list, and its line says what it serves and what a fix saves. A page whose largest paint is text says so. Fix the LCP image first: sizes="auto" helps lazy images, and the LCP image is eager.

CDN auto quality

A CDN's auto quality (Cloudinary q_auto) is set by the CDN, not by the URL, so an explicit quality is an extra step. ImgTune leaves it alone when its estimated saving is below a tenth of the group's served bytes: the hint says the CDN handles the quality, and that saving is not counted. Above a tenth, the hint names the explicit quality it would need. A CDN that negotiates the format (f_auto, imgix auto=format) gets no format change.

qualityNA

When an image's quality cannot be scored, usually because its original is unknown, it is counted as qualityNA. Its resolution still counts. Its estimate is then a floor.

Groups

Images are grouped by the key ImgTune builds for each element, so one layout is reported once. For each group ImgTune reports its estimated saving in bytes per page view, the sizes in use, a suggested sizes, and up to two sample URLs. The scan keeps the ten largest groups in the JSON and the top five in the Markdown summary.

Robots, blocks and what was not measured

A page that is disallowed by robots.txt, a page behind a bot wall, and a page that ends on another site are recorded with their status and are not measured. See Scanning any site.