What it removes is a class of human error
The value of an image CDN is not primarily the edge caching; it is that correctness stops depending on whoever uploads the file.
Without one, serving the right image requires someone to resize before upload, convert to a modern format, generate the variants, and reference the right one in the right context — every time, across everyone who touches the site. In practice that discipline fails, which is why oversized images are the most common performance finding on almost every platform.
With one, a designer uploading a 4000-pixel photograph still results in an 800-pixel AVIF being delivered to a phone, because the transformation happens per request based on the URL parameters and the browser's Accept header.