According to DEV Community, a web agency audited its own live site and found that the usual advice to add srcset and sizes did not fit as neatly as expected. Srcset is the markup that lets a browser pick between several sizes of one image. The author fetched 21 URLs from the sitemap with a plain Node script and recorded each image's markup, real dimensions and delivered size. The site had no responsive markup at all.
The first finding was that the homepage was not the heavy page. It has 61 images totalling 830 KB, but 59 are lazy-loaded, so they do not slow first paint. A service page was worse: its gallery loads six PNGs straight away. The author's advice is to count eager bytes, meaning those loaded before any scrolling, rather than total bytes.
The second finding came from arithmetic. On a 390 px phone with a pixel ratio of 3, a gallery card needs 1,098 device pixels across, and the source file is 1,024. So phones were slightly under-served, not over-served. On a 1440 px desktop at ratio 1, the card needs 413 and receives 1,024, which is about 6.1 times the pixels. Srcset helps the desktop case only.
The biggest gain came from changing format. Re-encoding the five visible gallery PNGs as JPEG at quality 80 cut 4,080 KB to 592 KB, about 85.5 percent less, with no change in pixel size. Resizing the PNGs made some files larger, for example 1,944 KB rising to 2,487 KB, because lossless compression handles resampled photographic detail badly.
The author lists limits. There is no field data, only lab measurements from one machine. Visual quality at q80 was not checked. WebP and AVIF were not tested, and the PNG growth came from one encoder's resize filter.
Why it matters
Image advice is often applied as a reflex. This audit shows that the same file can be too large on one device and too small on another, and that the cheapest fix may be a different file format rather than new markup. Fixing the wrong page or the wrong device wastes effort.
The PNG result is a useful warning. Making markup more correct can still make a page heavier if the image variants are bigger than the original. The figures come from one site and one tool, so treat them as an example of what to check, not a rule.
Source: DEV Community. This brief summarises their reporting; read the original for the full detail.