Post-training method that refines feed-forward 3D reconstruction models (Pi3, DA3, VGGT) on unlabeled video.
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.
PURE (Refining Feed-forward 3D Reconstruction with Unlabeled Post-training) is a research project that improves existing feed-forward 3D reconstruction models in dynamic scenes. It post-trains backbones such as Pi3, Depth Anything 3 (DA3) and VGGT on ordinary unlabeled videos, so no ground-truth 3D annotations are needed for those videos.
Training cues are derived from the videos themselves: monocular depth, dynamic masks and cross-frame matches. These supervise depth ordering, local geometry and static cross-frame consistency. Given a video or image sequence, the refined models output depth, camera poses and a colored point cloud.
The project page offers a "PURE Studio" demo for uploading a video or image sequence, along with example scenes across the three backbones. The page says code and model weights are provided, though no repository links were in the material reviewed. The authors note limitations in their long-sequence demo: some seams and moving-rider ghosting remain, and the accumulation is offline rather than live causal reconstruction. Authors are affiliated with several institutions; the vendor is not stated on the page.
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 builds 3D scenes from video or images. Existing models were improved using unlabeled ordinary videos converted to depth and motion cues. Open source with training scripts.
Closest neighbours by meaning, not just by category.
Research
Research method that reconstructs complete 4D (3D plus time) scenes from a single monocular video using Gaussian Splatting.
Model · Nvidia
Feed-forward model that repairs lifted 2D tracks and depth into accurate dense 3D trajectories from monocular video.
Model
Feed-forward 3D Gaussian Splatting model that reconstructs scenes from arbitrary input views, with or without camera poses.
Model
Feed-forward 3D scene reconstruction model that outputs triangle primitives directly usable as meshes for simulation.
Model
Feed-forward system that turns an RGB image or video into editable, textured per-object 3D meshes in under a minute.
Research · Preferred Networks
Generative method from Preferred Networks that reconstructs high-resolution 3D scene meshes from posed multi-view images.
