NVIDIA's 117M-parameter looped transformer for multi-view 3D reconstruction of camera poses and dense geometry.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
Déjà View is a research model from NVIDIA and academic collaborators that reconstructs camera poses and dense geometry (per-view depth and ray maps) from any number of input views. Instead of stacking many unique transformer layers, it applies a single transformer block recurrently to per-view features, with the number of refinement steps K exposed as an inference-time compute setting.
Per-view features come from a pretrained DINOv2 encoder. The looped block contains frame and global attention sub-blocks, and each application is conditioned on a continuous time interval. K is sampled per batch during training, so one checkpoint covers a range of step counts. Two lightweight heads decode the final state into depth and ray maps.
At 117M parameters, the authors report that it matches or beats larger feed-forward baselines (π³ and Depth Anything 3 Giant) on five benchmarks covering indoor, outdoor, object-centric and driving scenes. They report 8–10× fewer parameters and 1.9–2.3× less compute. A YouTube summary says code has been released, but no repository link appears in the provided material. The model is a research release rather than a hosted product.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
3D reconstruction model that turns multiple images into 3D Gaussian splats with camera poses by reusing one transformer block repeatedly. At 117M parameters it matches Depth Anything 3, which is about 10x larger; code released.
Closest neighbours by meaning, not just by category.
Model · Nvidia
NVIDIA research model that turns a single image into a queryable 3D Gaussian scene with semantics and object instances.
Model
Feed-forward 3D Gaussian Splatting model that reconstructs scenes from arbitrary input views, with or without camera poses.
Model
3D reconstruction model that recovers geometry and camera poses from photos or video using a loop transformer.
Research
Post-training method that refines feed-forward 3D reconstruction models (Pi3, DA3, VGGT) on unlabeled video.
Model · Nvidia
Feed-forward model that repairs lifted 2D tracks and depth into accurate dense 3D trajectories from monocular video.
Model · NVIDIA
NVIDIA research system where a code game engine drives a two-step streaming MiniMax-H3 video model for real-time rendering.