Take a look in NIP-94. It adds a hash to verify ifnthe picture has be doctored by the server, a blurhash to show when loading and a description for accessibility. It's a great spec.

#[4]​ is replacing urls with notes that you have to fetch to get the url which adds extra latency to image loading. I would rather just have this info in the kind1 tags so I don’t have to n+x notes for every note in the timeline.

this situation is exactly what major clients need to take into account, each breaking change imposes a tax on everyone else that is running a small client nostr: metions are superior to #[], but now I am forced to go back and update a bunch of clients I wrote just to regain what was previously working it's ok for new features, but breaking functionality must be embraced very slowly