{"slug": "deja-view-looping-transformers-for-multi-view-3d-reconstruction", "title": "Déjà View: Looping Transformers for Multi-View 3D Reconstruction", "summary": "NVIDIA researchers introduced Déjà View, a 117M-parameter multi-view 3D reconstruction model that applies a single transformer block recurrently instead of stacking unique layers, and it outperforms much larger feed-forward baselines on five reconstruction benchmarks spanning indoor, outdoor, object-centric and driving scenes while using 8–10× fewer parameters and 1.9–2.3× less compute than π³ and Depth Anything 3 — G. Because the number of refinement steps is sampled per batch during training, one checkpoint exposes step count as an inference-time compute knob. The team reports that the looped block also beats an otherwise identical variant with independent per-step parameters under matched training data and compute, which the authors say suggests explicit iteration is a stronger inductive bias for multi-view 3D reconstruction rather than just a cheaper substitute for capacity.", "body_md": "## TL;DR\n\n**Déjà View** reconstructs camera poses and dense geometry from any number of views by\napplying the **same transformer block** in a looped fashion. A single checkpoint\ncovers a range of step counts, so users can dial compute up or down at inference time.\nAt only **117M** parameters, Déjà View beats much larger\nfeed-forward baselines while using **8–10× fewer parameters and 1.9–2.3× less\ncompute** (π³, Depth Anything 3 — G) on five reconstruction benchmarks spanning\nindoor and outdoor scenes.\n\n## Abstract\n\nRecent feed-forward 3D reconstruction transformers have scaled to over a billion parameters, following the broader trend of increasing model capacity in computer vision. Yet emerging evidence suggests that contiguous transformer layers often behave like repeated applications of similar operations, and multi-view reconstruction transformers refine their predictions progressively across decoder depth. We posit that model depth partially buys iteration, paid for inefficiently in unique parameters, and instead make that iteration explicit in architecture.\n\nOur model, **Déjà View**, applies a single looped transformer block recurrently to\nper-view features for  refinement steps. Trained once, it exposes  as an\ninference-time compute knob, matching or outperforming substantially larger\nfeed-forward baselines across five reconstruction benchmarks spanning indoor, outdoor,\nobject-centric, and driving scenes, while using a fraction of their parameters and\ncomparable or lower compute. Importantly, the same looped block formulation\noutperforms an otherwise identical variant with **independent** per-step parameters\nunder matched training data and compute, suggesting that explicit iteration is not\nmerely a compute-efficient substitute for capacity but a stronger inductive bias for\nmulti-view 3D reconstruction.\n\n## Interactive viewer\n\nSlide the iteration control inside the viewer to watch **the same forward pass**\nrefine itself. Frustums are colored by per-camera pose error after Sim(3) alignment to\nground-truth poses.\n\n## Method\n\n**Déjà View** initializes per-view features from a pretrained DINOv2 encoder and\napplies a **single** transformer block — with frame and global attention sub-blocks —\nrecurrently  times to refine the state, with each application conditioned on its\ncontinuous time interval . Because  is sampled per batch from\n during training, one trained checkpoint covers any\nstep count in that range at inference. Two lightweight heads then decode the final\nstate into per-view depth and ray maps.\n\n## Results\n\nAt **117M** parameters, **Déjà View** leads average inlier ratio and pose AUC@30°\nacross all five benchmarks at the smallest parameter count of any baseline.\nBubble area is proportional to parameter count.\n\n## Qualitative comparison\n\nLeft viewer shows **Déjà View**. Use the dropdown on the right to compare against\nrecent feed-forward baselines on the same sequence. Pick a different example using\nthe strip below. All sequences are visualized discarding the bottom 25% of points\nby predicted confidence (if available).\n\n## Citation\n\n## Acknowledgements\n\nWe thank our colleagues at NVIDIA for valuable discussions and feedback.", "url": "https://wpnews.pro/news/deja-view-looping-transformers-for-multi-view-3d-reconstruction", "canonical_source": "https://research.nvidia.com/labs/dvl/projects/dvlt/", "published_at": "2026-11-13 20:35:23+00:00", "updated_at": "2026-11-13 20:50:16.591402+00:00", "lang": "en", "topics": ["computer-vision", "artificial-intelligence", "machine-learning", "ai-research"], "entities": ["NVIDIA", "Déjà View", "DINOv2", "Depth Anything 3", "π³"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/deja-view-looping-transformers-for-multi-view-3d-reconstruction", "markdown": "https://wpnews.pro/news/deja-view-looping-transformers-for-multi-view-3d-reconstruction.md", "text": "https://wpnews.pro/news/deja-view-looping-transformers-for-multi-view-3d-reconstruction.txt", "jsonld": "https://wpnews.pro/news/deja-view-looping-transformers-for-multi-view-3d-reconstruction.jsonld"}}