Breaking

German-led team unveils Seen2Scene to complete 3D scenes from partial real-world scans

June 29, 2026 at 05:28 EDT

Media Generation
Research & Papers

A research team led by the Technical University of Munich has presented Seen2Scene, a method that generates coherent, complete 3D scenes from incomplete real-world scans. Accepted to ECCV 2026, its code and paper are publicly available.

ECCV 2026 · TUM × University of Virginia

Seen2Scene: completing real 3D scenes from broken scans

A new method fills the gaps in incomplete LiDAR and RGB-D scans — reconstructing coherent, realistic 3D spaces by training directly on real incomplete data, not synthetic perfect scenes.

2026

Accepted to ECCV

Conditioning inputs: scans, layout boxes, text

Real-world datasets: ScanNet++, ARKitScenes

The shift: trained on real incomplete scans

Prior methods (SG-NN, NKSR) learned from synthetic complete data — and struggled in cluttered real rooms. Seen2Scene learns from messy reality.

Prior methods

Synthetic complete data

Seen2Scene

Real incomplete scans

How it reconstructs a scene

Partial scan

TSDF on sparse grid

→

Visibility mask

Hides unobserved regions

→

Flow matching

Sparse transformer fills gaps

→

Complete 3D scene

Coherent & realistic

Why it excites

Works directly with real-world data
High-fidelity completion from incomplete scans
Applications: AR/VR, digital twins, robot navigation, CG/VFX, driving sim

Still open

Research stage — limited large-scale use
Sparse-transformer compute cost untested
Stability on large scenes & real-time performance unproven

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.