CVPR 2026

PatchScene: Patch-based Voxel Diffusion for Large-Scale Scene Completion

Local diffusion. Global coherence. Infinite LiDAR scene completion.

Qingdong Xu^1*, Jiajun Zhu^1,2*, Shilin Zhu^4*, Xinjing He⁵, Chao Lu², Huanran Wang², Jiyao Zhang^3†

¹MEGVII Technology ²Qianli Technology ³Peking University

⁴Northeastern University, China ⁵Northwest Polytechnical University, Xi'an

^*Equal contribution ^†Corresponding author

PatchScene expands local voxel diffusion patches outward to complete large LiDAR scenes. — PatchScene completes large LiDAR scenes by repeatedly generating local voxel patches, fusing them into a coherent scene, and propagating reliable geometry outward.

Explicit Voxel Patch Diffusion

Instead of predicting continuous XYZ offsets for individual points, PatchScene denoises binary occupancy states inside high-resolution local voxel patches.

Spatio-Temporal Patch Fusion

Overlapping patches and adjacent frames are coupled during denoising, reducing patch boundaries, scattered artifacts, and frame-to-frame flicker.

Annular-Flow Completion

Generation follows the radial density pattern of LiDAR: high-confidence near-range structure guides sparse far-range regions, enabling spatially scalable completion.

Abstract

We propose PatchScene, a diffusion-based framework for large-scale LiDAR scene completion. Unlike existing methods that rely on global latent representations or dense voxel grids, PatchScene adopts a patch-based voxel diffusion paradigm that explicitly generates fine-grained geometry within localized 3D regions. To ensure coherent reconstruction at both spatial and temporal scales, we introduce a confidence-guided spatio-temporal fusion mechanism that integrates overlapping patches and adjacent frames in a unified generative process. Furthermore, Annular-Flow diffusion leverages the radial density pattern of LiDAR scans to progressively propagate high-fidelity information from near-range to far-range regions, enabling spatially unbounded scene completion.

Why Voxel Diffusion?

Outdoor LiDAR completion is not only a sparsity problem. The representation itself determines how difficult it is to generate precise geometry at scene scale.

Point-based diffusion

Precise XYZ offsets are difficult to recover

Methods such as LiDiff and ScoreLiDAR operate directly on points. Their diffusion process must move noisy points to the correct positions by predicting offsets along three continuous XYZ axes. Small errors easily accumulate into scattered points, smoothed surfaces, and distorted boundaries.

Explicit voxel diffusion

Occupancy turns generation into a regular prediction problem

Voxel space discretizes irregular point clouds into a structured grid. Instead of regressing precise continuous coordinates, the model predicts whether each voxel is occupied, making fine-grained geometry easier to reconstruct accurately.

The remaining challenge

Dense voxel grids are too expensive for outdoor scenes

High-resolution voxelization grows cubically with scene size. PatchScene resolves this bottleneck by dividing the global voxel grid into overlapping local patches and performing diffusion inside each compact patch.

Method: Divide, Denoise, Fuse, and Flow Outward

PatchScene makes large-scale completion tractable without giving up fine geometry: it solves local explicit voxel generation problems, then turns those local solutions into one temporally coherent scene.

PatchScene method overview with patch-based voxel diffusion, spatio-temporal fusion, and annular-flow completion.

Step 1 · Patch-based voxel diffusion

Diffuse occupancy inside local voxel patches

PatchScene decomposes the outdoor scene into overlapping 20 m by 20 m voxel patches. Each local patch is denoised independently in explicit occupancy space, allowing the model to generate fine geometry at a resolution that would be impractical for a single global dense grid.

Step 2 · Spatial consistency fusion

Fuse neighboring patches into a coherent scene

Independent patch generation can leave discontinuities in overlapping regions. During inverse diffusion, PatchScene projects neighboring noise predictions into a global field and stochastically couples 50% of overlap voxels with global context. This aligns adjacent patches while avoiding the blur introduced by deterministic averaging.

Step 3 · Annular-flow completion

LiDAR physics determines the generation order

Near-range LiDAR observations are dense and reliable, while distant regions are sparse and heavily occluded. Annular-Flow starts from the inner ring and expands outward, letting high-fidelity completed structure guide the next sparse ring at every denoising stage.

Step-by-step visualization of PatchScene expanding completed voxel patches from the center toward outer regions.

Benchmark Results

PatchScene consistently improves geometric accuracy, distribution fidelity, and volumetric completeness on both SemanticKITTI and KITTI-360.

SemanticKITTI

0.319CD ↓ 0.444JSD 3D ↓ 0.371JSD BEV ↓ 45.3IoU 0.5 ↑

Method	CD ↓	JSD 3D ↓	JSD BEV ↓	Voxel IoU ↑
Method	CD ↓	JSD 3D ↓	JSD BEV ↓	0.5	0.2	0.1
LMSCNet	0.641	-	0.431	30.8	12.1	3.7
LODE	1.029	-	0.451	33.8	16.4	5.0
MID	0.503	-	0.470	31.6	22.7	13.1
PVD	1.256	-	0.498	15.9	4.0	0.6
LiDiff	0.434	0.564	0.444	31.5	16.8	4.7
LiDiff(refine)	0.376	0.573	0.416	32.4	23.0	13.4
LiDPM	0.446	0.532	0.440	34.1	19.5	6.3
LiDPM(refine)	0.377	0.542	0.403	36.6	25.8	14.9
ScoreLiDAR†	0.412	0.589	0.425	29.7	13.0	3.6
ScoreLiDAR(refine)†	0.342	0.590	0.399	32.0	19.9	9.4
PatchScene	0.319	0.444	0.371	45.3	38.2	19.7

KITTI-360

0.356CD ↓ 0.424JSD 3D ↓ 0.341JSD BEV ↓ 51.8IoU 0.5 ↑

Method	CD ↓	JSD 3D ↓	JSD BEV ↓	Voxel IoU ↑
Method	CD ↓	JSD 3D ↓	JSD BEV ↓	0.5	0.2	0.1
LMSCNet	0.979	-	0.496	26.17	9.21	2.88
LODE	1.565	-	0.483	33.06	15.24	4.68
MID	0.637	-	0.476	33.05	21.32	11.30
LiDiff	0.564	-	0.459	33.23	17.55	4.88
LiDiff(refine)	0.517	-	0.446	33.43	22.04	11.84
ScoreLiDAR	0.472	-	0.444	-	-	-
ScoreLiDAR(refine)	0.452	-	0.437	-	-	-
PatchScene	0.356	0.424	0.341	51.8	42.5	21.3

SemanticKITTI baselines, metrics, and ground truth follow LiDPM; † denotes independently reproduced and evaluated results. KITTI-360 results are reported in the supplementary material.

Qualitative Comparison

Prior methods often remain sparse after completion or introduce hallucinated points during refinement. PatchScene produces denser surfaces with fewer boundary artifacts.

KITTI-360 qualitative comparison between LiDiff, LiDPM, ScoreLiDAR, input, ground truth, and PatchScene.

Temporal Consistency

PatchScene caches denoising predictions from the previous frame, aligns them to the current observation, and blends them with density-adaptive weights. This suppresses inter-frame flicker while improving geometry.

0.155 → 0.086RMSE, t0 to t1

0.159 → 0.081RMSE, t1 to t0

0.319 → 0.309CD with temporal fusion

PatchScene

PatchScene + Temporal Fusion

Spatial Scalability

Annular-Flow gives PatchScene a natural path to larger scenes: a model trained on 20 m LiDAR ranges can be directly applied to 50 m completion without retraining.

PatchScene trained on 20 meter range generalizes to 50 meter range completion. — 20 m training range to 50 m inference range: completed scenes preserve road surfaces, object boundaries, and global coherence in both open and narrow environments.

Comparison Videos

Side-by-side sequences make the same pattern visible over time: PatchScene is denser, cleaner, and more stable than point-offset or score-based baselines.

LiDiff

ScoreLiDAR

PatchScene

PatchScene + Temporal

Real-World Proprietary Dataset

Beyond SemanticKITTI, the supplementary material evaluates PatchScene on a high-fidelity real-world LiDAR dataset with 550 clips and more than 40 scene categories.

Scene category distribution and representative dense point clouds from the proprietary dataset.

PatchScene completion results on proprietary real-world LiDAR scenes.

PatchScene reports CD = 0.199 and Voxel IoU = 0.595 at 0.5 m on this dataset, showing robustness under complex real-world driving scenarios.

Citation

@inproceedings{patchscene2026,
  title     = {PatchScene: Patch-based Voxel Diffusion for Large-Scale Scene Completion},
  author    = {Xu, Qingdong and Zhu, Jiajun and Zhu, Shilin and He, Xinjing and Lu, Chao and Wang, Huanran and Zhang, Jiyao},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2026}
}