In bug 1829603, we cached the src attribute of images, but only for images without alt text. The thinking was that images with alt text are accessible, so the src isn't necessary there. However, it turns out that there is a fair amount of useless alt text out there, so further analysis of the image (e.g. using AI) can be necessary to make them reasonably accessible. To facilitate this, some of these analysis tools require the URL of the image. This won't always work - e.g. if authentication or other session data is required to fetch the image - but it may work in many cases.

Some analysis on my own Firefox session suggests this wouldn't consume as much memory as I feared, perhaps 1 to 2% extra a11y cache memory at maximum. Chromium already does this for all images. Finally, our current behaviour is a regression as compared with the old, non-caching architecture.

See this thread for the discussion that catalysed this, which includes a lot of additional context and real world examples:
https://nvda-addons.groups.io/g/nvda-addons/topic/99398978

Unfortunately, authors providing useless alt text is becoming increasingly common, so in the real world, we can't assume that having alt text makes an image accessible.

https://hg.mozilla.org/mozilla-central/rev/aa31e8faae1f