LaS-Comp: Zero-shot 3D Completion with Latent-Spatial Consistency
Authors: Weilong Yan, Haipeng Li, Hao Xu, Nianjin Ye, Yihao Ai, Shuaicheng Liu, Jingyu Hu
Organizations: National University of Singapore · University of Electronic Science and Technology of China · The Chinese University of Hong Kong · Changhong Intelligent Robot
This paper introduces LaS-Comp, a zero-shot and category-agnostic approach that leverages the rich geometric priors of 3D foundation models to enable 3D shape completion across diverse types of partial observations. Our contributions are threefold: First, \ourname{} harnesses these powerful generative priors for completion through a complementary two-stage design: (i) an explicit replacement stage that preserves the partial observation geometry to ensure faithful completion; and (ii) an implicit refinement stage ensures seamless boundaries between the observed and synthesized regions. Second, our framework is training-free and compatible with different 3D foundation models. Third, we introduce Omni-Comp, a comprehensive benchmark combining real-world and synthetic data with diverse and challenging partial patterns, enabling a more thorough and realistic evaluation. Both quantitative and qualitative experiments demonstrate that our approach outperforms previous state-of-the-art approaches. Our code and data will be available at LaS-Comp.
Figures & tables
Figure 1: Our new framework supports category-agnostic shape completion across diverse partial patterns, including (a) random crops, (b) single-view scans, and (c) missing semantic parts. It further supports both unconditional and text-guided completion, offering flexible control for real-world applications, see (d).
Figure 2: Overview of the LaS-Comp framework. Starting from Gaussian noise, the process iteratively refines a latent feature xt under the guidance of the partial input Sp . At each iteration t , this refinement is performed in two stages: the Explicit Replacement Stage (ERS) and the Implicit Alignment Stage (IAS) . The ERS explicitly injects the known geometry of Sp into xt to produce an updated latent xt∗ . The IAS then refines xt∗ using a gradient-based optimization, yielding a spatially aligned latent xt−dt for the next step. After the final iteration, the completed shape Sc is obtained by decoding the refined latent.
Figure 3: Overview of the Explicit Replacement Stage (ERS). At each timestep t , ERS decomposes the latent generation into two parallel branches. The clean branch (top) enforces spatial consistency, yielding x0∣t∗ . Concurrently, the noisy branch (bottom) enhances fidelity, producing x1∣t∗ . These two branch outputs are then interpolated to compute the final aligned latent xt∗ .
CD ↓ / EMD ↓
Table
Exe-Chair
Out-Chair
Old-Chair
Vase
Off-Can
Vespa
Tricycle
Trash
Couch
Average
SVDFormer [ 89 ]
5.48/6.68
3.20/5.85
0.79 /1.45
3.79/5.95
5.70/6.98
5.13/6.71
3.10/4.91
2.73/4.96
3.67/5.05
1.86/2.92
3.54/5.15
AdaPoinTr [ 80 ]
5.02/6.25
2.58/4.80
0.82 / 1.37
3.62/5.64
5.14/6.50
4.47/6.35
1.96/3.54
1.83 /3.66
1.21 /3.08
1.01 / 2.09
2.77/4.33
Shape-Inv [ 82 ]
1.58/2.84
3.59/5.75
1.36/2.12
4.55/7.39
3.91/7.40
3.10/4.77
4.36/7.24
5.05/7.40
2.48/3.83
1.95/2.76
3.19/5.15
P2C [ 14 ]
1.57/2.64
3.87/6.42
1.28/2.10
3.72/5.82
4.54/7.04
3.36/4.82
6.75/10.7
7.78/11.2
3.28/4.21
2.55/3.16
3.87/5.81
SDS-Comp [ 37 ]
1.35/2.30
1.96/2.65
2.51/3.92
2.77/3.77
3.00/5.25
3.79/4.28
3.36/5.73
3.18/3.49
2.69/3.21
2.95/4.56
2.74/3.93
PCDreamer [ 65 ]
0.82 /2.36
1.43/3.56
1.19/2.07
2.61/5.34
2.41 / 3.64
2.62 /3.89
2.20/4.30
2.53/3.82
1.55/2.47
1.47/2.61
1.88/3.41
Table 1: Quantitative comparisons on Redwood [ 10 ] . We highlight the best and second-best results.
Figure 4: Qualitative comparison on Redwood dataset [ 10 ] . We compare with various supervised and unsupervised methods [ 89 , 80 , 37 , 65 , 31 ] , and visualize the output as meshes utilizing the commonly-used mesh reconstruction method [ 54 ] .
CD ↓ / EMD ↓
Horse
Max-Planck
Armadillo
Cow
Homer
Teapot
Bunny
Nefertiti
Bimba
Ogre
Lucy
Dragon
Average
SVDFormer [ 89 ]
4.33/5.17
8.84/7.95
4.97/6.13
3.50/4.39
2.30/3.20
5.00/5.91
9.61/9.19
5.46/5.94
7.54/7.18
4.90/5.55
2.08/2.93
2.88/4.35
5.12/5.66
AdaPoinTr [ 80 ]
4.88/5.45
8.60/8.52
5.14/5.97
3.48/4.51
2.28/3.30
3.92/4.53
9.33/8.86
5.54/6.16
8.16/7.62
4.53/5.41
1.85/2.79
3.07/4.54
5.07/5.64
Shape-Inv [ 82 ]
6.55/10.4
4.94/5.72
4.79/7.22
4.74/7.83
2.36/3.75
3.29/4.53
5.79/6.58
4.39/5.70
5.29/6.85
5.76/9.51
2.99/4.07
4.48/7.91
4.62/6.68
PCDreamer [ 65 ]
2.52/4.34
5.20/3.85
2.78/4.53
1.79/3.06
1.54/2.97
1.95/3.57
5.48/5.66
2.72/3.43
3.83/5.39
2.57/3.97
1.50/2.61
2.41/4.97
2.86/4.11
ComPC [ 31 ]
1.29/1.77
1.22/ 1.51
2.18/3.23
1.78/1.90
1.32/1.65
1.09/1.34
1.46 / 1.76
1.84/2.13
1.65/1.90
1.56/2.11
1.97/2.70
1.99/3.10
1.61 /2.09
Ours (Direct3D-S2)
0.96 / 1.42
1.11 /1.63
1.15 / 1.67
1.21 / 1.63
0.87 / 1.26
0.88 / 1.24
1.55 /1.97
0.96 / 1.40
1.11 / 1.58
1.03 / 1.51
0.96 / 1.43
1.51 / 2.12
1.11 / 1.57
Table 2: Quantitative comparisons on the synthetic data [ 31 , 44 ] . We highlight the best and second-best results.
UCD ↓ / UHD ↓
ScanNet-Chair
ScanNet-Table
KITTI-Car
SVDFormer [ 89 ]
1.4 / 2.4
1.4 / 2.4
1.8 / 5.0
AdaPoinTr [ 80 ]
1.4 / 2.4
1.3 / 2.4
1.6 / 4.9
Shape-Inv [ 82 ]
4.0 / 9.3
6.6 / 11.0
5.3 / 12.5
P2C [ 14 ]
3.4 / 5.0
2.4 / 4.9
3.9 / 5.8
PCDreamer [ 65 ]
1.5 / 3.5
1.3 / 3.7
2.7 / 6.6
ComPC [ 31 ]
2.0 / 5.3
3.0 / 7.0
1.1 / 5.7
Table 3: Completion fidelity on ScanNet [ 15 ] and KITTI [ 21 ] .
CD ↓ / EMD ↓
Single Scan
Random Crop
Semantic Part
SVDFormer [ 89 ]
5.32 / 6.84
5.35 / 6.73
5.66 / 7.19
AdaPoinTr [ 80 ]
5.27 / 6.87
5.33 / 6.68
5.59 / 7.23
Shape-Inv [ 82 ]
4.98 / 7.31
5.39 / 6.79
5.37 / 6.91
SDS-Comp [ 37 ]
5.42 / 6.27
5.12 / 6.35
5.60 / 6.75
PCDreamer [ 65 ]
3.32 / 4.74
3.78 / 4.55
3.88 / 4.98
ComPC [ 31 ]
4.24 / 4.61
5.48 / 5.95
6.37 / 6.18
Table 4: Quantitative comparisons on our proposed Omni-Comp.
Figure 5: Visual examples on ScanNet [ 15 ] and KITTI [ 21 ] real-world datasets, which only contain real scans of table, chair, and car, with very sparse points.
Figure 6: Qualitative comparisons on cases with different partial patterns from our Omni-Comp benchmark. Our approach produces more reasonable results than the latest methods [ 31 , 65 ] .
MMD ↓ / TMD ↑
Redwood
Synthetic
PCDreamer [ 65 ]
1.88 / 0.13
2.86 / 0.11
ComPC [ 31 ]
2.01 / 0.72
1.65 / 0.62
Ours (Direct3D-S2)
1.81 / 1.35
1.23 / 0.89
Ours (TRELLIS)
1.62 / 0.99
1.20 / 0.94
Table 5: Completion diversity evaluation on Redwood [ 10 ] and synthetic data [ 31 , 44 ] .
CD ↓ / EMD ↓
Redwood
Synthetic
(a) Baseline (latent replacement)
2.15 / 2.93
2.33 / 2.78
(b) Full w/o ERS
3.42 / 4.94
3.53 / 4.85
(c) Full w/o PNS
1.94 / 2.56
2.27 / 2.67
(d) Full w/o IAS
1.88 / 2.14
1.17 / 1.56
(e) IAS w/ 10 optimization steps
1.42 / 1.87
1.20 / 1.48
(f) Full pipeline (Ours)
1.42 / 1.84
1.11 / 1.41
Table 6: Ablation studies. The baseline uses only latent replacement for completion.
Figure 7: Visual examples of the completion diversity from our method on the real-world data.
Figure 8: Visual comparison of the ablation studies. The red boxes highlight the artifacts and holes.
Figure 9: Illustration of the latent gap between the partial input Sp and the ground truth Sgt . Although Sp and Sgt share the same surface geometry in the observed regions, their latents obtained by the VAE encoder in the corresponding regions show an obvious difference.
Redwood
YCB
Synthetic
ID
Category
ID
Category
ID
Category
1
Bicycle
11
Cracker Box
21
Dinosaur
2
Plant
12
Mustard Bottle
22
Glass
3
Trash Bin
13
Banana
23
Guitar
4
Car
14
Bleach Cleanser
24
Headphone
5
Motorcycle
15
Skillet
25
Laptop
Table 7: Category list in the Omni-Comp benchmark. Representative object categories from Redwood [ 10 ] , YCB [ 4 ] , and Synthetic [ 63 ] .
Figure 10: Visual examples of the completion results under extremely noisy partial inputs. Despite the severe noise that heavily corrupts the observed points, our method can still recover a reasonable global structure and overall object silhouette, but many fine details and thin structures are degraded or missing, revealing the limitation of our model when strong noise overwhelms the underlying geometry.
Figure 11: Visual examples of the proposed Omni-Comp benchmark. We show the ground truth mesh, with its corresponding partial samples of random crop, single scan, and semantic part.
Figure 12: Visual examples of the proposed Omni-Comp benchmark. We show the ground truth mesh, with its corresponding partial samples of random crop, single scan, and semantic part.
Figure 13: Visual Comparison on the Redwood dataset [ 10 ] , with point cloud representation. Obviously, our method significantly outperforms prior approaches in most of the examples, with very high-quality geometry and good completion correctness.
Figure 14: Visual Comparison on the synthetic dataset [ 31 ] , with point cloud representation. Obviously, our method significantly outperforms prior approaches in most of the examples, with very high-quality geometry and good completion correctness.
Figure 15: More visual examples on the proposed Omni-Comp benchmark, under the random crop partial pattern. Our method consistently provides better completion results compared with the latest methods [ 65 , 31 ] , regardless of the partial pattern and object category.
Figure 16: More visual examples on the proposed Omni-Comp benchmark, under the single scan partial pattern. Our method consistently provides better completion results compared with the latest methods [ 65 , 31 ] , regardless of the partial pattern and object category.
Figure 17: More visual examples on the proposed Omni-Comp benchmark, under the semantic part partial pattern. Our method consistently provides better completion results compared with the latest methods [ 65 , 31 ] , regardless of the partial pattern and object category.
Figure 18: More visual examples to show the completion diversity, under unconditional and text-guided completion. In both completion settings, our method can provide reasonable results with good geometry.
3D shape completion from partial scans remains challenging for unseen categories and noisy real-world observations, where geometry alone is often insufficient for inferring missing structure. We present DinoComplete, a deterministic and efficient shape completion framework that augments geometric reconstruction with voxel-aligned semantic priors distilled from DINO features. First, we construct multi-view DINO feature volumes aligned with ShapeNet data and train a student network to predict dense semantic features directly from incomplete shapes. These predicted features capture global structure and part-aware semantic context while remaining aligned with the underlying geometry. We then integrate these distilled features into a completion network, where geometric and semantic voxel representations are fused through voxel state-space modeling. To enable efficient long-range reasoning without sacrificing resolution, we introduce a multi-scale voxel Mamba module that refines the fused features by combining full-grid and chunk-wise sequence modeling. Experiments on unseen ShapeNet categories and ScanNet objects show that DinoComplete achieves stronger completion quality than prior deterministic and generative based completion methods while using fewer parameters, requiring lower memory, and achieving faster inference. Our results demonstrate that distilling semantic priors from visual foundation models improves generalization and robustness in 3D shape completion.
Furkan Mert Algan, Eckehard Steinbach
Chair of Media Technology · Munich Institute of Robotics and Machine Intelligence · School of Computation Information and Technology, Technical University of Munich
Point cloud completion aims to infer a complete 3D shape from a partial point cloud and serves as a fundamental building block for downstream tasks such as reconstruction, editing, and simulation. Despite the recent progress, existing learning-based methods often implicitly rely on access to the ground-truth shape scale (GT-scale) during both training- and testing-time normalization, assuming privileged information that is unavailable in real-world inference. This hidden assumption limits practical deployment and can lead to severe completion artifacts, e.g., over- or under-completion and nested shells, once the oracle GT-scale cue is removed. We observe that the recent foundation image generation models exhibit a strong capability of understanding objects and geometries, and producing multi-view consistent renderings, making them promising priors for GT-scale-free 3D completion. Motivated by this insight, we propose ScaleBlind, a novel framework that leverages foundation-model-based image completion to recover global scale directly from partial inputs and then faithfully produces the 3D completion. Specifically, ScaleBlind dreams out complete multi-view appearances from rendered partial views, lifts the inferred missing regions back into 3D to obtain a geometry-aware coarse completion, and further refines it via a powerful cross-modal fusion network with the original partial point cloud. By harnessing 2D foundation priors, our method eliminates the need for accessing GT-scale information at inference. Moreover, it provides a principled bridge between 2D generative priors and 3D point cloud completion. Extensive experiments demonstrate the superiority of our framework, making ScaleBlind the new state-of-the-art for the point cloud completion task.
Shenghui Wu, Chen Wang, Yuan Feng +3
Shandong University, China · University of Edinburgh, United Kingdom
Recent 3D generative models produce high-quality geometry from a single image using large-scale priors and diffusion architectures. However, they assume complete visibility and single-view inputs, limiting applicability in multi-view, occluded, or editing scenarios. Although prior works address these challenges individually, they lack a unified framework for controllable 3D completion under diverse conditioning signals. We present Axolotl3D, a multi-modal and occlusion-aware 3D generation model that jointly conditions on images, visibility masks, camera parameters, and a partial point cloud. The point cloud serves as a geometric anchor promoting faithful shape completion, while camera parameters ensure consistent multi-view alignment in a shared 3D coordinate system. A unified training strategy synthesizes diverse conditioning regimes from large-scale 3D data, enabling robust cross-modal reasoning. Experiments on Toys4K and OmniObject3D demonstrate state-of-the-art performance under both clean and occluded settings, as well as strong results in real-world reconstruction and geometry-consistent editing.