NVIDIA Open-Sources Déjà View: 117M-Parameter Transformer for 3D Reconstruction

NVIDIA's NeurIPS 2026 paper Déjà View open-sources a 117M-parameter model that recycles a single transformer block to turn multi-view photos into 3D geometry and camera poses, matching models ten times its size; an interactive demo is available on Hugging Face.

2026-11-12 ~ 2026-11-14 · 3 related posts

• NVIDIA's Déjà View: 117M looped transformer matches models 10x larger at 3D reconstruction — NVIDIAAI · 2026-11-12
• NVIDIA open-sources Déjà View: a 117M model that loops one transformer block to match 10x larger models on 3D — CSProfKGD · 2026-11-13
• NVIDIA's Déjà View: 117M model loops one transformer block for 3D from photos — NVIDIAAI · 2026-11-14