Holi-Spatial 360

From sparse panoramas to spatial worlds

Holi-Spatial 360.

Generative Reconstruction from Sparse Panoramas
for Spatial Rendering and Intelligence

arXiv Code
RECONSTRUCTRENDERUNDERSTAND
Holi-Spatial 360 overview: sparse panoramic inputs, multi-room reconstruction, trajectory densification, 3D object detection and grounding, rendered depth, and spatial reasoning.
Sparse panoramas are reconstructed into geometrically multi-room 3DGS, densified with image and video generative priors, and converted into RGB-D, 3D grounding, and spatial-reasoning supervision

Framework

Open full-resolution figure
Holi-Spatial 360 framework: DA3 correspondence anchors for 3DGS optimization, generative trajectory densification with adapted DiFix and Wan-VACE, and verified spatial data curation.
DA3 correspondences anchor 3DGS optimization; adapted DiFix and Wan-VACE（1.3B） densify trajectories; verified RGB-D groundings yield 3D instances, captions, and spatial QA.

01 / Visual results

Explore four cases along matched camera paths.
Choose a baseline and compare in motion.

Sparse observations

Input panoramas

Two example references

Opening clip · Scene 00028
Input panorama 16 from Realsee scene 00028
Panorama 16
Input panorama 00 from Realsee scene 00028
Panorama 00
Holi-Spatial 360Ours
3DGSBaseline
Loading videos…
0:000:09
Three indoor scenes. Cuts between the 3-second clips are intentional.Open our video
About these comparisons Qualitative results · matched cameras · different native inputs

These videos follow the same camera paths within each case. Native inputs, reference budgets, and training conditions differ across methods, so these are qualitative demonstrations rather than an equal-input benchmark.

Baselines. Released DiFix and ArtiFixer operate on the independent, real-RGB-only 3DGS result. PanoWorld uses its own panorama-based reconstruction. GSFixer also operates on the independent 3DGS result.

The two panorama examples show representative source observations, not the full reference set for each method. Ziroom examples are the supplied full-panorama reprojections of the original cube faces.

02 / Spatial understanding

3D bounding boxes & depth.

RGB + 3D boxesHoli-Spatial 360
Depth0.15–8 m
0:000:14

Paired RGB and depth views along the same camera path. Depth is rendered from the underlying Gaussian geometry.