Do 3D reconstruction transformers really need a billion parameters, or are most of those layers just doing the same thing over and over?

Introducing Déjà View: a single transformer block, looped K times, that matches or beats models 8–10× its size with lower compute. 🧵