NVIDIA Research has released Déjà View (DVLT), an open 3D reconstruction model that turns unordered photographs or video into 3D geometry, depth maps, and estimated camera positions.
And it uses just 117 million parameters.
That's approximately 8–10 times smaller than several leading 3D reconstruction transformers, including VGGT, Pi3, and Depth Anything 3 Giant.
Developed by researchers from NVIDIA, ETH Zurich, the University of Toronto, and the University of Modena and Reggio Emilia, Déjà View introduces a different approach to multi-view 3D reconstruction.
Instead of stacking dozens of unique transformer layers, it repeatedly applies the same shared transformer block.
The model improves its reconstruction through multiple refinement steps, without requiring a separate set of weights for each iteration.
WHAT MAKES DÉJÀ VIEW DIFFERENT?
• 117M parameters: a compact architecture compared with many billion-parameter 3D vision models.
• Unposed image reconstruction: accepts unordered RGB photographs or video frames without requiring known camera positions.
• 3D geometry estimation: predicts dense depth maps, per-pixel rays, and 3D point clouds.
• Camera reconstruction: estimates camera intrinsics and extrinsics directly from visual input.
• Looping Transformer: reuses a shared attention block instead of relying on a deep stack of independent layers.
• Adjustable inference: control the number of refinement iterations K to balance computational cost and reconstruction quality.
• No retraining required: the released checkpoint supports different iteration counts, from K=8 to K=16.
• Competitive benchmarks: NVIDIA reports matching or outperforming substantially larger models across five 3D reconstruction benchmarks.
• Efficient computation: the researchers report 1.9–2.3x lower compute than selected large baselines.
• Open implementation: training code, inference scripts, evaluation tools, and pretrained weights are available.
• Interactive visualization: explore reconstructed point clouds, depth maps, and camera trajectories using Rerun and the included Gradio demo.
WHY DOES THIS MATTER?
Modern AI 3D reconstruction methods can require billions of parameters and substantial GPU resources.
Déjà View challenges the assumption that better reconstruction necessarily requires increasingly large models.
Instead, it uses repeated computation with shared weights to refine the geometry.
Potential applications include:
• 3D scene reconstruction from photographs
• Robotics and visual navigation
• AR and VR environment mapping
• Camera tracking and Structure-from-Motion
• 3D Gaussian Splatting and NeRF initialization
• Digital twins and spatial computing
• Research into efficient vision transformers
There are important limitations.
Déjà View produces depth, camera estimates, and reconstructed point clouds, not automatically textured, production-ready 3D meshes.
Its reported advantages depend on the benchmark and configuration. The model does not guarantee real-time performance or perfect geometric accuracy.
The source code is primarily Apache 2.0, while the pretrained model weights are distributed under NVIDIA's non-commercial research license.
Still, the architecture points toward an interesting direction for AI development:
What if smarter iteration matters more than adding billions of parameters?
NVIDIA Research project:
https://research.nvidia.com/labs/dvl/projects/dvlt/
GitHub source code:
https://github.com/nv-tlabs/dvlt
Model weights:
https://huggingface.co/nvidia/dvlt
Research paper:
https://arxiv.org/abs/2605.30215
Would you choose a smaller iterative model over a billion-parameter 3D reconstruction network?
#NVIDIAResearch #3DReconstruction #ComputerVision
YouTube / ABVX
