Breaking

3D-RE-GEN turns a single photo into editable 3D indoor scenes

June 29, 2026 at 18:37 EDT

Media Generation
Research & Papers
Open Source

A research team at the University of Tübingen has released 3D-RE-GEN, a method that generates editable 3D indoor scenes—with individual objects separated from the background—from a single photograph. The paper was posted to arXiv on December 19, 2025, with code released under the MIT license.

Dec 19, 2025 · University of Tübingen

3D-RE-GEN: A Full, Editable 3D Room From a Single Photo

An open-source (MIT) framework that reconstructs each object as a textured, swappable 3D mesh — separated from the background and grounded on the floor in one pass.

4-DoF

Constrained transform: x/z move, yaw, scale — locked to the ground plane

A-Q

Application-Querying: two-panel prompt fills in occluded object parts

534★

GitHub stars under MIT license

Ablation: removing each component degrades accuracy

Both novel components measurably improve the result when present.

Without 4-DoF constraint → worse geometry

0.51

IoU
no 4-DoF

0.68

F-Score
no 4-DoF

Chamfer Distance rises to 0.030.

Without A-Q → worse appearance

0.27

SSIM
no A-Q

0.66

LPIPS
no A-Q

Higher LPIPS = larger perceptual error.

The compositional pipeline

Single photo
input image

→

Segment
Grounded-SAM instances

→

2D→3D
Hunyuan3D assets

→

Layout
4-DoF grounded scene

Strengths

Objects sit stably on the floor — no floating or intersecting
Each object is an editable, swappable mesh
Generates backgrounds, unlike SAM3D
Validated on synthetic, real-world and outdoor scenes

Caveats

Research project — no pricing, limited real-world reports
Setup needed: SAM weights, mamba/venv environment
Dependencies like Hunyuan3D-2.0 carry their own licenses
Some are research/non-commercial only — verify before deploying

Continue reading

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