I have mixed feelings about this.

From a obvious point of view it feels absolutely bad spamming relays with fat notes, but if the relay is my one, and I'm thinking about the Gossip model here, it could make absolutely sense to have a unique source of truth for notes and their assets, that uses a unique protocol. Images are an important chunk within an information piece, so I would not give them to a generic third part.

A client could also offer its users a relay/web server/CDN layer to improve the images reachability and the download performances.

There is a proposed NIP-95 that is quite similar: