Back to feed

NVIDIA Unveils Compact Déjà View Model for 3D Reconstruction

Researchers from NVIDIA, ETH Zurich, and the University of Toronto have presented the open-source Déjà View model for reconstructing scenes from unordered photographs.

The architecture addresses Structure-from-Motion and dense 3D reconstruction tasks, simultaneously recovering geometry, depth maps, and camera positions.

The key difference lies in using a recurrent cyclic transformer block with shared weights instead of a deep stack of unique layers.

This approach reduced the model's parameter count to 117 million, making it 8–10 times more compact than similar models like VGGT and Depth Anything 3.

Despite its small size, the system outperforms heavier neural networks on ScanNet++ benchmarks and demonstrates high accuracy on indoor and outdoor scenes.

The cyclic structure allows dynamically adjusting the number of inference steps K without retraining, balancing streaming tracking speed with sub-millimeter detail.

Key solution characteristics:

  • Compactness: only 117 million parameters
  • Efficiency: recurrent transformer with shared weights
  • Functionality: simultaneous prediction of ray map, depth, and camera poses
  • Flexibility: compute control at inference time
  • Open source: code and weights available under a free license

The full source code, preprocessing pipeline, and Rerun visualization scripts are published on GitHub. Model weights are hosted on Hugging Face and ready to run via the Accelerate library in BF16 mode.

6.4K views

More from this channel Metavershche and AIshche

Similar in this category Technology