Jonathan Lorraine#1131
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. 🧵
Traditional 3D reconstruction pipelines like COLMAP operate in a loop, growing the scene piece-by-piece. 🧩🔄
With DejaView, we introduce this inductive bias for feed-forward 3D reconstruction—running a single alternating-attention block in a loop! Awesome work from the team 👇
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. 🧵
Jonathan Lorraine#1131Do 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. 🧵
