If I understand correctly, the events contain text and are stored and sent by relays, and so are less censorable because those events are available via many relays. The non-text media forms are sent as links in the events or links in the text and clients load them from their respective URLs. I cannot find a citation but I think this was a deliberate design decision because much problematic content is in images and videos, so by not directly relaying images and videos it lowers the risk (as well as resource requirements) for those who run relays. A lower burden to run a relay means more relays, more relays means a lower chance of censorship.

Can anyone point us to the link for the why? Maybe I'll find it later.