HIVE-3D: Hierarchical Voxel Enhancement for High-Quality 3D Scene Generation

ICML 2026

Bin Zang*, Wenting Zheng*, Xiaoliang Luo, Zhiyuan Fang, Shi Li, Lvchun Wang, Wei Yu, Yi Zhao, Tian Xie, Yuchi Huo, Rengan Xie^†

¹ State Key Laboratory of CAD&CG, Zhejiang University

² China Mobile (Jiangxi) Virtual Reality Technology Co., Ltd.

* Equal contribution · † Corresponding author

▤ Paper arXiv⌘ Code · Soon◇ Model · Soon

HIVE-3D progressively enhances a scene from coarse to fine hierarchy levels — HIVE-3D progressively enhances a coarse scene into high-resolution components and a coherent final scene.Click to expand ↗

Abstract

High-quality scenes,
one hierarchy at a time.

Recently, a line of works can generate impressive 3D objects from a single image, but they are limited by restricted representation resolution, making them unsuitable for 3D scene generation. We introduce HIVE-3D, a novel method for high-quality 3D scene generation based on a hierarchical voxel enhancement framework. Given a single scene image, HIVE-3D first produces a coarse initial scene, aligns hierarchical 2D image components with 3D scene components, and organizes them into a component tree. A voxel super-resolution model then refines each target instance while maintaining consistency with its coarse voxels. Coarse-to-fine enhancement of images and voxels produces high-resolution, high-quality 3D scenes and achieves state-of-the-art performance.

Method

From one image to a structured, high-resolution 3D scene.

Instead of forcing the entire scene through one restricted voxel grid, HIVE-3D decomposes enhancement into a coarse-to-fine hierarchy.

Overview of the HIVE-3D method — Method overview: scene-tree construction, voxel super-resolution, and hierarchical scene generation.Click to expand ↗

2D-to-3D Scene Tree Construction

Align hierarchical 2D image components with the initial 3D scene and organize the matched components into a tree.

Voxel Super-Resolution

Refine each target instance with its fine image crop and coarse voxels, preserving structure while recovering detail.

Hierarchical Scene Generation

Enhance the scene from root to leaves, then register the refined components back into one coherent 3D scene.

Results

Fine geometry returns without sacrificing the whole scene.

Across synthetic and real images, the hierarchy restores small objects, thin structures, and local texture while retaining global coherence.

Qualitative comparison with prior 3D scene generation methods — Qualitative comparison with MIDI, Gen3DSR, SceneGen, TRELLIS, and HIVE-3D.Click to expand ↗

Additional HIVE-3D generation results — Additional generation results.Click to expand ↗

Voxel super-resolution ablation.Click to expand ↗

Geometry · Full model

0.0035CD ↓84.34F-Score ↑0.7449IoU ↑

Image-space · Ours

0.79SSIM ↑13.39PSNR ↑0.33LPIPS ↓0.97CLIP ↑

Method	SSIM ↑	PSNR ↑	LPIPS ↓	CLIP ↑
Gen3DSR	0.76	10.96	0.34	0.91
MIDI	0.80	12.85	0.25	0.92
SceneGen	0.79	12.25	0.31	0.96
TRELLIS	0.80	13.31	0.31	0.95
Ours	0.79	13.39	0.33	0.97

Limitations

Where the current pipeline can improve.

HIVE-3D relies on the initial scene produced by TRELLIS. If objects are missed or incomplete at this stage, correspondence and registration may become unreliable.

Citation

BibTeX

@inproceedings{zang2026hive3d,
  title     = {HIVE-3D: Hierarchical Voxel Enhancement for High-Quality 3D Scene Generation},
  author    = {Zang, Bin and Zheng, Wenting and Luo, Xiaoliang and Fang, Zhiyuan and Li, Shi and Wang, Lvchun and Yu, Wei and Zhao, Yi and Xie, Tian and Huo, Yuchi and Xie, Rengan},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning},
  series    = {Proceedings of Machine Learning Research},
  volume    = {306},
  year      = {2026}
}