Pure turns ordinary videos into full 3D worlds

Pure is a new open-source 3D reconstruction project that turns ordinary videos or photo sequences into full 3D environments, without needing a single frame of expensive 3D ground-truth data.

How it actually works

Instead of training a giant model from scratch on scarce, lab-collected 3D scans, Pure takes existing feed-forward reconstruction backbones like Pi3, DA3, and VGGT, and refines them through unlabeled post-training.

The researchers fed regular, unannotated videos into the model. Pure extracts complementary geometric cues directly from the video frames, including monocular depth, dynamic masks, and cross-frame feature matches. It uses these visual clues to teach itself depth ordering and local geometry in dynamic scenes without requiring ground-truth 3D supervision.

In one demo, Pure processes a 2,000-frame motorcycle ride broken into 128-frame clips. As the bike rides through the environment, an overhead 3D point cloud builds out synchronously in the corner of the screen, mapping out the whole trajectory without point subsampling.

Specs, setup, and trade-offs

The entire release is surprisingly lightweight at just 13.2GB. The team open-sourced the codebase, the pre-trained weights, and the full training scripts. You can head over to their project page, hit the Code button, and follow the setup guide to run it locally on your own machine.

That said, it does have some real constraints. The reconstruction is an offline temporal accumulation rather than live causal reconstruction, meaning you cannot run it real-time on a live camera stream just yet.

Dynamic subjects can also produce ghosting artifacts. In the motorcycle demo, the team noted that visible seams and rider ghosting remain, with 7 of 15 seams meeting their recorded geometry thresholds. It is not ready for pixel-perfect VFX work out of the box.

Why it matters

High-quality 3D data has long been one of the nastiest bottlenecks in spatial AI. Pure proves that raw, unannotated video can act as effective post-training fuel to dramatically sharpen dynamic geometry.

For roboticists, indie game devs, and spatial computing tinkerers, this is an immediate win. Clocking in at 13.2GB with training scripts in hand, you can clone it locally and start fine-tuning on your own specialized footage without spending tens of thousands of dollars on 3D annotations.