A new vision method turns one photo into a 3D scene mesh, a surface model of the place shown, while estimating surfaces the camera never saw. In an Oct. 6 preprint, researchers adapt an object-focused 3D generator for indoor and outdoor scenes; they report better geometry scores than seven competing methods on scanned scenes and preference in at least 73% of blind comparisons on 121 in-the-wild images.
A scene covers more space than a single object, so squeezing it into one fixed-size 3D volume can erase nearby detail or leave distant structures out. The team adapted Trellis 2, a pretrained generator for 3D objects, by dividing a scene into chunks whose physical size grows with distance: small chunks keep detail near the camera, while larger ones reach farther across a room or street.
The method also tells the generator which parts of 3D space are open, visible in the photo or hidden behind a surface. It ties image features to the visible geometry, then generates chunks in sequence and aligns their overlapping regions. To address the scarcity of outdoor training data, the team added about 4,000 synthetic outdoor scenes to existing indoor datasets.
Against seven single-image reconstruction methods on seven Tanks and Temples scenes, the paper reports a 24% lower normalized Chamfer distance, meaning a smaller mismatch between generated and reference surfaces, and a 30% higher F1 score, meaning more generated surface points matched the reference within 10 centimeters while accounting for missed reference points. Each figure is against the strongest baseline on that metric. On 32 ScanNet++ scans, its geometry scores also edged the strongest baseline. For 121 in-the-wild photos, where ground-truth meshes were unavailable, participants preferred the method in at least 73% of blind pairwise comparisons with each baseline. Those comparisons measure visual preference, not geometric accuracy.
In an ablation using the same trained model, adaptive chunks took five minutes versus 15 minutes for fixed three-meter chunks and also scored better on reconstruction. That timing applies to the paper’s test setup.
The output is geometry without textures. It depends on estimated depth and camera intrinsics, the calibration values used to map pixels into 3D, and the authors say errors can carry from one generated chunk into later ones. The meshes are not guaranteed to be watertight, so holes may need repair.
Sources
- “Building Rome from a Single Image,” arXiv preprint, submitted Oct. 6, 2026
- Authors’ project page and results
- Microsoft Research: TRELLIS.2, the object-generation model adapted in the paper
- Figure from the authors’ preprint, used under CC BY 4.0.