VoxLution: Evolving Voxels for High-Fidelity 3D Scene Generation

Daoyi Gao1*, Xinyi Zhang1*, Alexey Bokhovkin1, Lei Li2, Angela Dai1
1Technical University of Munich   2University of Virginia
*Equal contribution
Paper arXiv Video Code (coming soon) BibTeX
VoxLution teaser: a generated ScanNet++-style scene refined from 25 cm to 1.5625 cm voxels

VoxLution generates 3D indoor scenes with a sparse autoregressive cascade that refines a coarse scene structure from 25 cm down to 1.5625 cm voxels, capturing sharp surfaces and fine geometric detail.

From Coarse Voxels to Fine Geometry

Abstract

High-resolution 3D indoor scene generation is crucial for immersive applications such as virtual reality, spatial computing, and robotics simulation. We introduce VoxLution, a coarse-to-fine transformer-based approach that generates high-fidelity indoor scenes with rich details, overcoming the overly smooth geometry typical of existing generative models. First, we generate a low-resolution spatial layout; then, based on this structural information, we progressively refine the scene geometry locally in chunks. Key to our approach is formulating 3D super-resolution as a sequence modeling problem over multi-resolution sparse voxel grids. To achieve this, we introduce a training-free, resolution-invariant tokenization scheme that compactly encodes the geometric information of eight octants into a single decimal number. We formulate the transformer sequence to include multiple resolutions, allowing the coarse layout to guide the generation of fine local details. To ensure coherence across local chunks, we incorporate adjacent chunk boundary information directly into the transformer sequencing. VoxLution achieves scene generation at a finer voxel resolution than existing baselines, demonstrating state-of-the-art performance on standard benchmarks with 59.6% lower FPD and 51.1% lower FID on 3D-FRONT, and 70.5% lower FPD and 73.8% lower FID on ScanNet++.

Video

Generated Scenes

Method

VoxLution pipeline

VoxLution synthesizes a scene as a multi-resolution autoregressive cascade. A coarse scene transformer first generates the coarse scene structure as a single-level token sequence, conditioned on the scene type. Cascade transformers then refine the scene chunk by chunk, each stage doubling the resolution, down to 1.5625 cm voxels.

Results

Real-world scans (ScanNet++)

Scenes generated by VoxLution from the ScanNet++ distribution
3D scene generation from the real scan distribution of ScanNet++, with fine geometric detail and coherent room-scale structure.
Comparison on ScanNet++
Comparison on ScanNet++. VoxLution generates fine geometric detail and realistic real-world clutter.

Synthetic scenes (3D-FRONT)

Comparison on 3D-FRONT, close-ups
Comparison on 3D-FRONT. VoxLution captures fine structures such as thin chair and table legs, cabinet dividers, and hanging clothes.
Comparison on 3D-FRONT, rooms
Comparison on 3D-FRONT. VoxLution recovers finer geometric detail and more coherent room-scale layouts.

Layout-conditioned generation

Layout conditioning
VoxLution is conditioned on a layout of open-vocabulary labeled bounding boxes, which enter every stage as layout tokens.
Layout-conditioned comparison
Layout-conditioned generation compared with Seen2Scene under the same input layouts.

BibTeX

@article{gao2026voxlution,
  title   = {VoxLution: Evolving Voxels for High-Fidelity 3D Scene Generation},
  author  = {Gao, Daoyi and Zhang, Xinyi and Bokhovkin, Alexey and Li, Lei and Dai, Angela},
  journal = {arXiv preprint arXiv:2610.41501},
  year    = {2026}
}

Acknowledgements

This project is funded by the ERC Starting Grant SpatialSem (101076253). We also acknowledge the EuroHPC Joint Undertaking for providing access to the EuroHPC supercomputers MareNostrum 5, hosted by the Barcelona Supercomputing Center (BSC), Spain, and LEONARDO, hosted by CINECA, Italy.