XNews

NVIDIA Unveils Déjà View Looped Transformer for Multi-View 3D Reconstruction

NVIDIA Unveils Déjà View Looped Transformer for Multi-View 3D Reconstruction

NVIDIA researchers introduced Déjà View, a 117M-parameter model that reconstructs camera poses and dense geometry from any number of views by applying a single transformer block recurrently. The model matches or outperforms feed-forward baselines up to 10x larger across five benchmarks while using 8–10x fewer parameters and 1.9–2.3x less compute, with the number of refinement steps exposed as an inference-time compute knob. The work was accepted as an oral paper at NeurIPS 2026, with paper and code released.

Event version 1.

Key insights

Single looped block replaces depth

Déjà View applies one transformer block recurrently to per-view features for K refinement steps, making iteration explicit in architecture rather than buying it through unique parameters. The same looped block outperforms an otherwise identical variant with independent per-step parameters under matched training data and compute.

Compute knob at inference

K is sampled per batch from a range during training, so a single checkpoint covers any step count in that range at inference. Users can dial compute up or down to fit their budget.

Small model, strong benchmarks

At 117M parameters, Déjà View leads average inlier ratio and pose AUC@30° across all five benchmarks at the smallest parameter count of any baseline. It uses 8–10x fewer parameters and 1.9–2.3x less compute than π³ and Depth Anything 3.

Method and outputs

Per-view features are initialized from a pretrained DINOv2 encoder, and two lightweight heads decode the final state into per-view depth and ray maps. Intermediate decoding shows progressively sharper geometry and more accurate camera poses across iterations.

Publication and availability

The work is an oral paper at NeurIPS 2026, with paper, code, an interactive viewer and BibTeX released.

Source posts