GeoWeaver turns ordinary long videos into drift-free 3D worlds

Turning a quick five-second clip into a neat 3D mesh is pretty easy for modern AI models. But try walking around your house recording for three minutes, and things fall apart fast.

The AI loses track of scale. The couch it measured in the living room suddenly shrinks when you return from the hallway, camera trajectories twist, and the final 3D scene turns into a smeared, drifting mess.

Researchers just unveiled GeoWeaver, a new framework engineered specifically to solve long-video reconstruction drift. Tested across multiple long-sequence benchmarks, it achieved the lowest average error among competing models.

Slicing footage and stitching it tight

Feeding an entire continuous minutes-long video directly into a single neural network blows up GPU memory. The common shortcut has been chunking videos into bite-sized segments, but that usually creates severe scale drift where the seams never line up.

GeoWeaver handles this with a two-part system that lives up to its name.

First, a Geometric Prior Model splits the video into small manageable chunks, predicting depth maps, confidence scores, and rough camera parameters for each segment to lock down local shapes.

Then comes the heavy lifting: Test-Time Adaptation. Instead of blindly gluing pieces together, GeoWeaver runs sequential initialization followed by global chunk-level alignment, refining camera poses, depth corrections, and lens intrinsics from coarse to fine.

By leveraging dense correspondences across adjacent frames, overlapping chunks, and long-range visual loops, a robust optimization objective balances 2D reprojection with 3D consistency. When the camera loops back around, the world stays anchored.

Ablation tests showed that this adaptation pipeline even improves trajectory accuracy when plugged into entirely different geometric backbones, proving the assembly math works on its own merits.

Why it matters

Taming long-sequence drift is the missing puzzle piece for accessible spatial computing.

If this framework eventually lands in everyday pipelines, creators, game designers, and indie VFX artists won't need bulky LiDAR rigs to capture real spaces. A casual phone walkthrough could be enough to generate clean, scale-accurate 3D environments without distorted geometry.

For the AI industry, it provides a clever detour around the brute-force compute trap. While many frontier labs try to shove massive video contexts straight into giant feed-forward networks, GeoWeaver proves that smart chunking paired with rigorous geometric optimization during inference can deliver top-tier accuracy at a fraction of the memory footprint.

That said, temper your expectations for now. The team has only published the research paper, with no open-source code or checkpoints released yet. Running test-time optimization also means you will spend extra compute during inference rather than getting an instant export. It is an impressive leap on paper, but we will have to wait for working code before knowing how it handles truly unconstrained handheld chaos.