WorksComputer Science
Any2Building: Unifying Point Clouds and Images for Building Reconstruction via Shared Spatial Queries
Abstract
Abstracting structured buildings from limited observations is essential for 3D urban modeling and digital twins. Partial point clouds and overhead images of the same building provide incomplete 3D surface measurements and visual cues to its footprint, roof geometry, and rooftop structures, respectively. This motivates a unified framework that aligns geometric and visual cues in 3D and reconstructs buildings from either or both modalities. Building on this connection, we present Any2Building, a unified model for Level of Detail 2 (LoD2) building reconstruction from partial point clouds, overhead images, or both, with optional coarse geometric priors. At its core is an occupancy-grounded shared spatial queries representation that organizes available observations within a common 3D grid. Supervised by complete building surface voxels, these queries connect observational evidence to target geometry, enabling a shared generator to learn structural priors across tasks and providing spatial guidance for recovering missing structure. A trained voxel conditional surface-generation and low-poly remeshing pipeline then converts the predicted structure into compact LoD2 building models. To support this framework, we introduce BuildingWorld-Image, comprising approximately one million paired samples of partial point clouds, overhead images, and LoD2 building models across 18 cities. Experiments on multiple benchmarks demonstrate that Any2Building outperforms existing methods in geometric reconstruction accuracy across unimodal and multimodal point-cloud completion, single-image building generation, and prior-assisted generation, while maintaining detailed and robust recovery when point cloud input modality is degraded.
ing generation, and prior-assisted generation, while maintaining detailed and robust recovery when point cloud input modality is degraded, enabling modality alignment and fusion as well as task unification to promote positive transfer across reconstruction tasks.
Introduction
Structured 3D building models support urban digital twins [2] and applications such as urban understanding [43] and visual localization [44]. Level of Detail 2 (LoD2) modeling [17] captures a building’s spatial organization in a compact, structured representation that preserves the geometry and spatial relationships of its roof sections and detailed rooftop structures. Building reconstruction often relies on limited observations, including partial point clouds [13, 15, 16, 22] and overhead images [18, 30]. Point clouds provide partial geometric measurements, while overhead images reveal appearance and layout but leave height and hidden surfaces ambiguous. Although these heterogeneous observations can originate from the same building, point cloud based building reconstruction and image-based building generation are still largely treated as separate tasks. This common origin motivates a unified reconstruction framework that recovers LoD2 building from either modality independently or their combination.
Recent building reconstruction methods have made progress by drawing on powerful 3D generation pipelines [1, 25, 31, 32], but their designs remain tailored to individual observation settings and are difficult to extend directly to complementary modalities. BuildAnyPoint [13] reconstructs structured buildings from partial point clouds, but its voxel-aligned conditioning does not readily incorporate image evidence as a spatially grounded 3D condition. In prior-assisted image-based building generation [18], extending noise-based prior injection to partial point clouds can obscure fine structural cues. Meanwhile, general-purpose 3D generators offer useful inspiration through their efforts to unify heterogeneous conditions within a shared framework [6, 11, 28, 42]. Axolotl3D [11] and Hunyuan3D-Omni [28] concatenate visual and geometric tokens for attention-based conditioning, but rely on the generator to resolve cross-modal structural relationships while inferring missing geometry—a challenging joint inference problem under incomplete observations.
This direct coupling is particularly challenging in building reconstruction, where low-resolution orthographic imagery provides limited height and facade cues, yet reveals roof layouts and footprint boundaries essential for LoD2 reconstruction. However, direct feature fusion does not readily align these visual cues with the fine-grained geometric evidence in partial point clouds, as illustrated by the direct-fusion example in Figure 1(b). The remaining challenge is to establish consistent geometric correspondence between heterogeneous observations and complete building structure, enabling modality alignment and fusion as well as task unification to promote positive transfer across reconstruction tasks.

Figure 1. Any2Building across heterogeneous observations. A single jointly trained model accepts (a) partial point clouds, (b) sparse points with overhead imagery, (c) imagery cropped by instance bounding boxes, and (d) imagery with a coarse geometric prior. The reconstructed geometry supports compact LoD2 modeling and downstream texture synthesis.
To address this challenge, we introduce Any2Building, a unified LoD2 building reconstruction model that supports point cloud input and remote sensing image inputs, together with height or box priors. Our main technical innovation is a set of shared queries initialized on a low-resolution 3D voxel grid. The available inputs are encoded as conditioning tokens, and shared spatial queries aggregate their features at common 3D locations. Supervision with complete building surface voxels encourages the queries to capture the overall building structure, including unobserved regions, before they condition the generator. By relating different observations to the same underlying geometry, the queries further guide the generator, enabling shared structural learning across point-cloud completion and image-based generation, together with optional coarse geometric priors. The generated structure latent is decoded into a voxel grid. We find that these voxels alone suffice for continuous mesh reconstruction by a trained surface generator, without additional point or image conditioning. Low-poly remeshing and texturing then produce compact Lod2 building assets.
To support learning across observation settings, we introduce BuildingWorld-Image, an image extension of BuildingWorld [17]. It comprises approximately one million aligned triplets of partial point clouds, overhead images, and LoD2 building models across 18 cities. These observations share the same geometric targets, enabling structural learning and evaluation across input modalities. Extensive experiments on multiple public datasets and BuildingWorld-Image dataset demonstrate that our method not only achieves superior performance on individual tasks but also benefits from complementary modalities when observations from point cloud modality are limited.
Our contributions are summarized as follows.
We propose Any2Building, a unified model for LoD2 building reconstruction from partial point clouds, overhead images, or both, with optional coarse geometric priors. The model combines complementary observations and promotes positive transfer across reconstruction tasks.
We design a spatial conditioning module based on shared 3D queries. Supervised by complete building voxels, these queries align heterogeneous observations with building geometry and provide spatial guidance for generator.
We introduce BuildingWorld-Image, a dataset containing approximately one million aligned triplets of partial point clouds, real overhead images, and LoD2 building models across 18 cities, supporting training and evaluation across observation settings.
Related Work
Point cloud completion for building reconstruction. General point cloud completion methods recover missing geometry [9] through global feature encoding [37, 40], hierarchical decoding [27, 33], voxelized representations [34], and geometry-aware transformers [3, 38, 39]. Multimodal methods, including EGIInet [36], DuInNet [21], and MAENet [20], further incorporate paired images [10]. These methods output point sets, requiring subsequent processing for structured mesh reconstruction.
To recover coherent architectural structure, PolyFit [23] and City3D [15] assemble planar primitives into compact meshes through constrained face selection. ArcPro [16] predicts architectural programs encoding building layers and their hierarchy, while Point2Building [22] autoregressively generates mesh vertices and faces from point clouds. For existing meshes, RoLoPM [4] combines offset surface extraction, simplification, and geometric flow to produce low-poly models while preserving visual fidelity. BuildAnyPoint [13] combines point-conditioned diffusion with autoregressive mesh generation, while UniCo [5] uses learnable primitive proxies to jointly predict point geometry and assembly-ready primitives non-autoregressively. Both rely on observed 3D geometry through voxelized point conditioning and shared point cloud features. Incorporating images therefore requires an additional mechanism to spatially ground 2D appearance.
Image-based building generation. Recent image-to-3D models, including TRELLIS, TRELLIS.2, and SAM 3D [1, 31, 32], provide strong generative priors for recovering geometry from visual observations. At the building level, BuildDiff [30] generates point clouds through single-image-conditioned diffusion, while SatSkylines [18] introduces geometric control by interpolating coarse-prior latents with Gaussian noise. However, extending this noise-space injection to partial point clouds can obscure fine geometric cues, while coupling geometric guidance with noise strength limits independent control over local structure. Related urban generation methods synthesize larger scenes through compositional modeling [19, 35] and satellite-conditioned generation [12, 14, 24, 29]. Overhead appearance nevertheless leaves height and hidden surfaces ambiguous, while coarse priors do not specify detailed architectural geometry.
Geometrically conditioned 3D generation. Geometric conditioning enables more explicit control over generated shapes. SDFusion [6] supports generation from partial geometry and images, while CLAY [42] accommodates diverse controls, including point clouds, voxels, and bounding boxes. SpaceControl [7] introduces spatial constraints at test time without additional training. Hunyuan3D-
Omni [28] and Axolotl3D [11] concatenate visual and geometric tokens for attention-based conditioning, leaving the generator to infer how these heterogeneous cues jointly constrain missing geometry. This calls for a shared representation that spatially organizes complementary observations and relates them to complete architectural structure across input settings.
Method
Any2Building generates LoD2 buildings from a partial point cloud , an overhead image , or both, with an optional coarse geometric prior . Its central component is a set of shared spatial queries that aggregate observations on a low-resolution 3D grid. Supervised by complete building occupancy, these queries form an intermediate representation of overall structure before conditioning the generator. As illustrated in Figure 2, the query features guide a flow-matching DiT through query-enhanced (QE) tokens and spatial prompts. The generated structure is decoded into a surface-voxel grid and converted into a compact LoD2 mesh. All geometric inputs, spatial queries, and output voxel grids share a normalized building coordinate system linked to the georeferenced overhead imagery. This allows generated buildings to be aligned with the source imagery and mapped back to their real-world locations, even in the image-only setting.
Figure 2. Any2Building overview. (a) Available observations and optional priors guide building generation through shared spatial queries, followed by surface generation and low-poly remeshing. (b) Queries on a coarse voxel grid aggregate observation features and optional priors. Voxel occupancy supervision encourages an intermediate representation of overall building structure, which conditions the DiT through QE tokens and spatial prompts.
Multimodal Input Tokenizers
We first encode the available observations for aggregation by shared spatial queries. Point clouds and images yield modality-specific tokens of width , while optional geometric priors provide descriptors for query initialization. The point encoder uses four EdgeConv stages with 16-nearest-neighbor graphs in coordinate space. Farthest-point sampling selects 2,048 and then 1,024 centers, repeating samples when necessary. The resulting center coordinates and local features give
where the two MLPs encode local geometry and spatial position, respectively, yielding . A frozen DINOv3 encoder [26] extracts image patch tokens . The token sequences retain their modality-specific organization; shared spatial queries associate them with 3D locations.
LoD0 priors are 3D building bounding boxes; LoD1 priors are building footprints extruded to their corresponding heights. Each prior is voxelized into a binary surface-occupancy grid , leaving interiors empty. For each voxel block corresponding to a cell of the query grid, we compute
where marks occupied voxels with at least one empty six-connected neighbor; locations outside the grid are treated as empty. An MLP encodes the descriptor , representing surface occupancy presence, boundary presence, and occupancy fraction. For thin surface shells, and may coincide.
Shared Spatial Query Conditioner
We aggregate available observations into an intermediate representation on a coarse 3D voxel grid, supervised by complete-building surface occupancy. The grid need represents both observed and missing regions, providing structural guidance before the generator.
We initialize queries at fixed grid-cell centers . Mean-pooled, normalized modality features and availability flags form a global descriptor , with missing modalities contributing zero features. The initial coarse queries are
where denotes concatenation. When a prior is available, its descriptors update the queries as
where scales the update. Without a prior, . All query locations are retained, allowing generation beyond the occupied prior voxels. Prior dropout during training reduces dependence on this additional input.
The queries aggregate available modalities through cross-attention:
Missing modalities yield zero readouts, and is broadcast across queries. Residual self-attention and feed-forward blocks, followed by normalization, produce the spatial-query features . An auxiliary head predicts surface occupancy with the loss
The targets are obtained by max-pooling the decoded reference grid of the complete building to . They represent surface occupancy rather than filled interiors and are shared across input settings. This auxiliary loss only supervises not the final voxels.
The query features guide generation in two ways. First, they enrich the base condition tokens through cross-attention. For a single observation modality, or . For point-image inputs, cross-attention updates the point tokens with image tokens and auxiliary completion features, preserving the point-token count and width. The completion features come from 512 learnable queries distinct from the spatial queries. The enriched condition is
Second, provides spatial prompts that retain their 3D grid indices.
DiT with Spatial Prompts
The flow-matching DiT generates a structure latent, conditioned on through cross-attention in every block. Selected blocks additionally receive aligned spatial prompts:
where is a zero-initialized, layer-specific linear projection and scales the residual. Shared grid indexing aligns each query with its corresponding latent location (Figure 3). Figure 3(b) visualizes the spatial prompt residuals injected into two DiT blocks, labeled Branch A and Branch B. The layer-specific projections provide spatial guidance that emphasizes different parts of the building.
Figure 3. Spatial prompt injection. (a) Zero-initialized projections inject aligned residuals. QE tokens condition every block. (b) A/B show RMS residual deviations from different spatial prompt block’s surface mean. shows their difference.
Given a complete-building latent , Gaussian noise , and time , the flow path and target velocity are
Here denotes the clean target latent, and specifies the residual noise level at . The training objective is
For voxel-level geometric supervision, clean latents are estimated from and the predicted velocity using the coefficients of Eq. (9). The estimated latents are decoded into voxel grids, on which applies BCE and Dice supervision.
Training samples four input modes, , , , and , which share the conditioner, generator, and complete-building target representation. QE tokens and spatial prompts are jointly dropped during training for classifier-free guidance, which is a common training setting for DiT.
At inference, we integrate the flow from Gaussian noise at to and denote the sampled endpoint by . The structure decoder maps this latent to a voxel grid.
LoD2 Construction
After obtaining the voxel grid , we generate local surface latents at its occupied voxel coordinates . We find that the predicted voxel geometry is sufficient for surface generation without additional image or input point-cloud features. We therefore remove image conditioning from the surface-stage structured-latent generator [32] and train it on building data using
where denotes reference surface latents, their interpolation with Gaussian noise , and the target flow velocity. Training uses reference voxel coordinates, while inference uses the predicted coordinates .
A pretrained shape decoder converts the generated surface latents into a mesh, and RoloLPM [4] produces a compact polygonal model retaining principal roof and facade geometry. This surface-construction pipeline is shared by all input modes. For texturing, we synthesize plausible facade references using ControlNet [41], conditioned on untextured mesh renderings and appearance descriptions inferred from the overhead image and city name, then use pretrained texture generator [32] to texture the low-poly mesh with these references. The resulting LoD2 buildings can be mapped back into their real-world scenes by reversing the normalization and applying the associated georeferencing transforms.
Experiments
Evaluation Protocol
Dataset. Our BuildingWorld-Image contains 1,011,196 real overhead-image–low-poly mesh pairs across 18 cities before splitting, with approximately one million pairs.
used for training. The test set contains 2,000 buildings from unseen cities and held-out regions within training cities. Coarse priors are constructed from bounding boxes or height-extruded footprints. Subsets with textured ground-truth meshes support appearance evaluation through matched four-view renderings, with Hong Kong examples shown in Figure 4. We also use Building-PCC [8] and SatSkylines [18] for point-cloud completion and prior-assisted generation comparisons, respectively. Dataset construction, splits, and evaluation protocols are detailed in Appendices A.

Figure 4. Real imagery with textured building ground truth in BuildingWorld-Image. Remote-sensing observations, source textured meshes, and instance-colored geometry.
Evaluation settings and metrics. Within each input setting, methods share the evaluation frame and input preprocessing. Building-PCC follows the metric definitions and reporting scale of BuildAnyPoint [13]. On BuildingWorld-Image, we report , , surface F-score at a normalized threshold of 0.02, Earth Mover’s Distance (EMD), uniformity (Uni.), and density-aware Chamfer distance (DCD). Both CD variants and EMD are scaled by . The main experimental tables report F-scores as percentages. PSNR and SSIM evaluate appearance on the textured-reference subset. Voxel-support studies use voxel F1. Detailed metric definitions and alignment procedures are provided in the Appendices B.
Training and inference. Our model is trained for 200k iterations on eight NVIDIA A800 GPUs. Inference uses the same GPU type, with 12 flow-sampling steps and a guidance scale of 3.0. Learned 3D generation baselines are fine-tuned on the corresponding training split for 200k optimizer updates per independently trained generative stage.
Point-Cloud Completion
Unimodal point-cloud completion. We first evaluate on Building-PCC [8], which includes LiDAR, SfM, and sparsely sampled point clouds. We compare Any2Building
with point completion baselines. Except for City3D [15], we train each method independently on this dataset. Table 1 shows that Any2Building achieves the best completion and low-poly accuracy, with PC and LP F-scores of 96.12% and 94.27%, respectively. It also has the smallest vertex and face count ratio errors, indicating close agreement with reference mesh complexity. Figure 5 shows recovered roof structures and coherent low-poly geometry from incomplete observations.

Figure 5. Completion and low-poly examples. Partial inputs, completed point sets, and compact geometric outputs are shown alongside the reference. These selected examples illustrate how point completion quality affects the recovered architectural form.
| Method | Point cloud (PC) CD ↓ | F-score (%) ↑ | EMD ↓ | Uni. ↓ | Low-poly mesh (LP) CD ↓ | F-score (%) ↑ | ( ↓) | ( ↓) |
| City3D [15] | - | - | - | - | 1.49 | 45.43 | 1.62 (0.81) | 1.47 (0.63) |
| Point2Building [22] | - | - | - | - | 1.20 | 52.92 | 1.29 (0.38) | 1.26 (0.41) |
| FoldingNet [37] | 1.47 | 41.27 | 0.29 | 2.43 | 1.89 | 35.14 | 1.42 (0.61) | 1.58 (0.74) |
| PCN [40] | 0.74 | 86.14 | 0.15 | 0.69 | 1.08 | 79.32 | 1.28 (0.39) | 1.36 (0.47) |
| TopNet [27] | 0.97 | 25.43 | 0.35 | 0.61 | 1.60 | 19.28 | 1.35 (0.48) | 1.41 (0.52) |
| GRNet [34] | 1.41 | 74.16 | 0.19 | 1.13 | 1.86 | 67.35 | 1.31 (0.44) | 1.33 (0.45) |
| SnowflakeNet [33] | 1.38 | 60.42 | 0.17 | 1.52 | 1.77 | 56.31 | 1.26 (0.36) | 1.29 (0.41) |
| PoinTr [8] | 0.41 | 85.09 | 0.12 | 0.25 | 0.55 | 77.46 | 1.18 (0.27) | 1.21 (0.32) |
| AnchorFormer [3] | 0.39 | 82.35 | 0.13 | 1.27 | 0.60 | 75.18 | 1.22 (0.31) | 1.24 (0.35) |
| AdaPoinTr [39] | 0.44 | 70.24 | 0.15 | 1.49 | 0.64 | 65.37 | 1.15 (0.25) | 1.19 (0.29) |
| UniCo [5] | 0.43 | 73.16 | 0.14 | 1.44 | 0.63 | 68.29 | 0.76 (0.26) | 0.72 (0.20) |
| BuildAnyPoint [13] | 0.35 | 91.38 | 0.10 | 0.04 | 0.66 | 80.14 | 0.83 (0.23) | 0.80 (0.27) |
| Ours | 0.20 | 96.12 | 0.07 | 0.03 | 0.30 | 94.27 | 1.05 (0.08) | 1.03 (0.07) |
Table 1. Point-cloud completion and low-poly reconstruction on Building-PCC. and are mean per-instance predicted-to-reference vertex and face count ratios. The parenthesized values and are their mean absolute deviations from 1.
We further evaluate on BuildingWorld-Image. As shown in Table 2(a), Any2Building achieves the best results at both the point-cloud and low-poly stages. Compared with AdaPoinTr, it reduces final low-poly from 11.05 to 8.65 and increases the F-score from 92.00% to 97.19%. The qualitative comparisons in Figure 6(a) show more complete roof structures and cleaner building boundaries, consistent with the improved geometric accuracy.
Figure 6. One model, four observation settings. Any2Building reconstructs buildings from (a) partial points, (b) images with partial points, (c) a single overhead image, and (d) images with a rough prior. The examples highlight roof articulation, coherent building massing, and the use of complementary evidence. All compared models that support training (with fire logo) were trained under the corresponding task-specific experimental settings.
| Point cloud (PC) | Low-poly mesh (LP) | |||||||||
| Method | Input | CDL ↓ | F-score (%) ↑ | CDL ↓ | CDL ↓ | F-score (%) ↑ | EMD ↓ | Uni. ↓ | DCD ↓ | |
| (a) Unimodal point-cloud completion | ||||||||||
| PoinTr [38] | P | 11.24 | 91.00 | 12.68 | 0.79 | 88.00 | 92.42 | 0.34 | 0.52 | |
| AnchorFormer [3] | P | 12.35 | 88.00 | 13.87 | 1.04 | 85.00 | 140.98 | 0.60 | 0.55 | |
| SnowflakeNet [33] | P | 10.42 | 93.00 | 11.69 | 0.78 | 91.00 | 80.68 | 0.08 | 0.47 | |
| AdaPoinTr [39] | P | 9.81 | 94.00 | 11.05 | 0.62 | 92.00 | 90.57 | 0.14 | 0.48 | |
| UniCo [5] | P | 10.35 | 93.00 | 11.28 | 0.79 | 91.00 | 88.94 | 0.10 | 0.47 | |
| BuildAnyPoint [13] | P | 11.47 | 93.00 | 12.88 | 1.83 | 92.00 | 86.56 | 0.08 | 0.46 | |
| Ours | P | 9.25 | 96.25 | 8.65 | 0.40 | 97.19 | 68.12 | 0.04 | 0.30 | |
| (b) Multimodal point-cloud completion | ||||||||||
| PoinTr [38] | P | 46.31 | 58.00 | 50.03 | 21.18 | 54.00 | 177.35 | 4.17 | 0.70 | |
| AnchorFormer [3] | P | 45.72 | 61.00 | 49.28 | 24.25 | 58.00 | 206.22 | 0.57 | 0.69 | |
| SnowflakeNet [33] | P | 41.66 | 60.00 | 45.27 | 14.91 | 57.00 | 219.88 | 0.95 | 0.68 | |
| AdaPoinTr [39] | P | 54.28 | 62.00 | 58.89 | 46.25 | 59.00 | 207.00 | 0.75 | 0.67 | |
| UniCo [5] | P | 53.61 | 50.00 | 58.42 | 43.51 | 55.00 | 172.48 | 0.12 | 0.71 | |
| BuildAnyPoint [13] | P | 52.94 | 40.00 | 57.20 | 20.64 | 58.00 | 166.13 | 0.08 | 0.80 | |
| EGIInet [36] | P+I | 41.35 | 55.00 | 44.98 | 13.45 | 51.00 | 283.07 | 1.60 | 0.95 | |
| DuInNet [21] | P+I | 49.82 | 50.00 | 53.76 | 17.67 | 46.00 | 294.32 | 1.55 | 0.95 | |
| Hunyuan3D-Omni [28] | P+I | 43.15 | 53.00 | 44.52 | 10.62 | 50.00 | 145.27 | 0.27 | 0.92 | |
| MAENet [20] | P+I | 42.47 | 60.00 | 46.11 | 11.08 | 52.00 | 134.51 | 0.20 | 0.94 | |
| Ours | P+I | 27.53 | 75.45 | 26.00 | 10.07 | 77.41 | 119.74 | 0.06 | 0.66 |
Table 2. Unimodal and multimodal point-cloud completion on BuildingWorld-Image. P denotes partial points and I denotes an overhead image. Group (b) uses sparser partial observations than group (a). PC outputs precede surface generation and low-poly reconstruction. LP denotes final low-poly meshes.
Multimodal point-cloud completion. Under sparser observations, all methods receive the same partial points, with paired images additionally provided to multimodal methods. Any2Building achieves the lowest low-poly of 26.00 and the highest F-score of 77.41% in Table 2(b). Figure 6(b) illustrates how image guidance helps recover missing building structures, including roof layouts and building components that are insufficiently constrained by sparse point observations. The paired study in Table 4(b) isolates the contribution of imagery under matched point observations.
Image-Based Building Generation
Single remote-sensing image. We compare with TRELLIS [31], TRELLIS.2 [32], and SAM 3D [1], all fine-tuned on BuildingWorld-Image. Any2Building reduces CDL from 54.89 to 38.90 and increases the F-score from 49.20% to 62.82% compared with TRELLIS (Table 3(a)). The full textured pipeline also achieves the highest PSNR (12.63 dB) and SSIM (0.34). Figure 6(c) illustrates the recovery of distinct building volumes and their spatial arrangement, including the separation of towers in the multi-volume example. We use bounding-box crops to retain building shadows and surrounding context and to avoid propagating errors from intermediate semantic segmentation.
| Geometry | Appearance | ||||||||
| Method | Input | CDL ↓ | CDL ↓ | F-score (%) ↑ | EMD ↓ | Uni. ↓ | DCD ↓ | PSNR (dB) ↑ | SSIM ↑ |
| (a) Single-image generation | |||||||||
| TRELLIS [31] | I | 54.89 | 9.14 | 49.20 | 111.08 | 0.03 | 0.66 | 10.91 | 0.26 |
| TRELLIS.2 [32] | I | 65.09 | 9.90 | 31.05 | 124.57 | 0.03 | 0.79 | 10.80 | 0.28 |
| SAM 3D [1] | I | 55.80 | 8.53 | 44.34 | 117.59 | 0.04 | 0.71 | 9.26 | 0.23 |
| Ours | I | 38.90 | 6.32 | 62.82 | 93.16 | 0.03 | 0.60 | 12.63 | 0.34 |
| (b) Prior-assisted generation | |||||||||
| SpaceControl [7] | I+R | 34.77 | 3.83 | 61.93 | 86.15 | 0.02 | 0.57 | 11.65 | 0.30 |
| Hunyuan3D-Omni [28] | I+R | 32.46 | 3.51 | 60.82 | 83.27 | 0.04 | 0.55 | 11.38 | 0.29 |
| SatSkylines [18] | I+R | 33.12 | 3.42 | 59.62 | 84.69 | 0.05 | 0.56 | 11.21 | 0.28 |
| Ours | I+R | 17.23 | 1.53 | 78.64 | 57.82 | 0.03 | 0.37 | 14.05 | 0.41 |
Table 3. Image-based building generation on BuildingWorld-Image. I denotes an overhead image and R denotes a coarse geometric prior. Backgrounds are retained in (a) and removed in (b). PSNR and SSIM compare matched four-view renderings of predicted and reference textured meshes.
Orthophoto with a geometric prior. Using the same orthophotos and coarse geometric priors, we compare Any2Building with SpaceControl [7], Hunyuan3D-Omni [28], and SatSkylines [18]. Any2Building achieves the lowest CDL of 17.23 and the highest F-score of 78.64% (Table 3(b)). The coarse priors constrain the overall building extent and massing but leave detailed roof geometry unspecified. Figure 6(d) illustrates how our method recovers roof structures beyond those represented by the coarse priors. Additional evaluations on 500 Skylines-50K assets show consistent advantages over the compared methods under both LoD0 and LoD1 priors (Appendix § 8).
| Variant | P | I | P+I | I+R |
| Spatial-query conditioning | ||||
| Condition concat | 91.86 | 53.34 | 60.71 | 57.32 |
| w/o occupancy supervision | 94.75 | 54.83 | 71.08 | 64.03 |
| w/o query read-back | 93.78 | 51.17 | 72.44 | 70.38 |
| w/o spatial prompts | 92.15 | 51.47 | 69.13 | 57.02 |
| Single-task training | 95.30 | 57.38 | 72.91 | 76.83 |
| Full Any2Building | 97.19 | 62.82 | 77.41 | 78.64 |
Table 4. Ablation of the unified generation pipeline. (a) Conditioning components and joint versus single-task training. (b) Paired P and P+I inference under point degradation. (c) Surface generation in the point-cloud-input setting. Both surface variants use predicted voxel coordinates, and one additionally receives point-cloud conditioning. F-scores are reported as percentages.
| Degradation | Points | P (Sparse) | P (Sparse)+I | F (pp) |
| None | 2048 | 74.39 | 77.41 | +3.02 |
| Random | 205 | 54.08 | 72.63 | +18.55 |
| Random | 41 | 38.67 | 64.32 | +25.65 |
| Spatial crop | 41 | 19.43 | 46.51 | +27.08 |
Table 4 (PDF p. 8). Image contribution: LP F@2%
Table 4 (PDF p. 8, second). Surface generation: final low-poly reconstruction
Ablation Study
We evaluate spatial-query conditioning, shared learning, image complementarity, and surface generation on the full 2,000-building BuildingWorld-Image test set. Geometric metrics are computed on low-poly meshes.
Spatial-query conditioning. Table 4(a) compares the full conditioning design with direct concatenation of the available input tokens, including prior tokens when present. We also remove occupancy supervision, query read-back, or spatial prompts individually and retrain each variant under the same training protocol. The full model achieves the highest F-score across all four settings. Without occupancy supervision, the image-only and prior-assisted F-scores decrease from 62.82% to 54.83% and from 78.64% to 64.03%, respectively, supporting the value of supervising the intermediate queries with complete building geometry. Removing either conditioning path also lowers performance. Without spatial prompts, the prior-assisted F-score drops to 57.02%. Figure 7(b) visualizes query occupancy

Figure 7. From spatial queries to compact geometry. (a) Surface generation reduces voxel staircase artifacts before low-poly remeshing. (b) Query and decoder occupancy alongside target surface voxels. (c) Predictions from single-task and jointly trained models. Quantitative component comparisons are reported in Table 4(a).
alongside the final decoder output.
Shared learning across tasks. We compare the joint model with one specialist per input setting, using the same number of training steps and the same inference inputs. Joint training samples the four input modes with equal probability and improves the final low-poly F-score by 1.89, 5.44, 4.50, and 1.81 percentage points for P, I, P+I, and I+R, respectively. The gains in P and I show that shared training also benefits reconstruction when only one observation modality is available at inference. Figure 7(c) compares predictions from single-task and jointly trained models. The jointly trained model predicts voxel structures that yield more regular low-poly meshes with higher structural fidelity, illustrating the benefits of shared learning across tasks.
Image contribution under sparse observations. We compare P and P+I inference using the same sparse points setting with imagery. With the original 2,048 point partial observations, adding imagery increases the F-score from 74.39% to 77.41%, a gain of 3.02 percentage points (Table 4(b)). With 41 retained points, adding imagery raises the F-score from 38.67% to 64.32% under random subsampling and from 19.43% to 46.51% after random spatial cropping followed by subsampling. Imagery thus provides greater benefits as geometric evidence becomes more sparse.
Surface generation and low-poly reconstruction. Surface refinement improves low-poly reconstruction with both PolyFit and RoLoPM (Table 4(c)). With RoLoPM, it increases the F-score from 95.88% to 97.19% and reduces from 9.83 to 8.65. Figure 7(a) illustrates the reduction in voxel staircase artifacts. Additional point-cloud conditioning slightly reduces accuracy and increases runtime. We therefore generate surfaces conditioned only on the predicted voxel coordinates.
Conclusion
We presented Any2Building, a unified model for LoD2 building reconstruction from partial point clouds, overhead images, or both. The model also supports image-based reconstruction with coarse geometric priors. Supervised by complete-building surface occupancy, our shared spatial queries aggregate input features on a common 3D grid and condition the same structure generator across input settings. We generate surfaces from the predicted voxels alone and remesh them into compact LoD2 models. We also introduced BuildingWorld-Image, comprising approximately one million aligned triplets of partial point clouds, overhead images, and LoD2 building meshes across 18 cities. Experiments show improved reconstruction accuracy across unimodal and multimodal point-cloud completion, single-image generation, and prior-assisted generation. Ablations show that joint training improves reconstruction from either modality alone, and that adding imagery improves completion from sparse or spatially incomplete point clouds.
Any2Building: Supplementary Material
This supplementary material provides dataset details, implementation and evaluation protocols, and additional analyses for Any2Building. Appendix A describes dataset construction, quality screening, splits, and image-based evaluation and rendering. Appendix B presents implementation, geometric metrics, ablation protocols, and surface-generation and low-poly reconstruction results. Appendix C reports encoder and query-resolution ablations and analyses of geometric priors, alongside qualitative component comparisons and textured visualizations.
BuildingWorld-Image: Construction and Evaluation
BuildingWorld-Image extends an existing geometric collection with geographically grounded visual observations. Its purpose is to connect building geometry, partial point measurements, and real overhead imagery within a common instance-level representation. This section details the data composition, construction and refinement procedures, partitioning, image-based evaluation protocols, and representative cases, including source textured meshes and instance-level correspondences (Figure S3).

Figure S3. Textured ground truth from neighborhood to instance. Each group shows the remote-sensing context and a selected building, with corresponding source textured mesh and instance-colored geometry. Geometry, camera, and framing are shared between the textured and colored mesh views. Source textures are retained; instance colors are display attributes, and road context is schematic. These are source references, not model predictions. Imagery © HKSAR Government / Lands Department · Roads © OpenStreetMap contributors
Composition and Geographic Coverage
Table S1 reports the tabulated BuildingWorld data pool before train–test partitioning. It contains 1,583,024 unique city-qualified geometric targets, including 1,011,196 targets paired with real overhead imagery after quality filtering. This tabulated image subset spans 18 cities on four continents and includes residential, industrial, and dense urban buildings. Perth contributes 3,843 image–mesh pairs. The image count is a subset of the geometry count, so the two columns are not additive. Counts refer to source building targets rather than augmented image views, point-cloud realizations, or prior variants. Some source targets contain grouped structures. The paired subset connects visual observations to geometry; the broader geometry pool provides additional architectural examples.
| City | Geometry | Image |
| Boston | 167,410 | 63,857 |
| Calgary | 322,649 | 224,190 |
| Cambridge | 17,041 | 11,191 |
| Cape Town | 274,598 | 125,202 |
| Edmonton | 370,673 | 246,540 |
| Greater Geelong | 886 | 596 |
| Hong Kong | 25,544 | 11,724 |
| Melbourne | 8,909 | 5,895 |
| Mississauga | 145,080 | 139,489 |
| Montreal | 24,866 | 24,652 |
| New York | 47,815 | 44,370 |
| Perth | 4,111 | 3,843 |
| Philadelphia | 2,907 | 2,117 |
| San Francisco | 91,359 | 65,583 |
| Tokyo | 34,507 | 10,998 |
| Toronto | 4,255 | 4,197 |
| Wellington | 15,213 | 13,165 |
| Yarra | 25,201 | 13,587 |
| Total | 1,583,024 | 1,011,196 |
Table S1. BuildingWorld data pool before partitioning. Geometry counts unique city-qualified low-poly targets. Image counts targets with a paired overhead image after quality filtering, without tier quotas. Image is a subset of Geometry; counts exclude augmentation multiplicity. These are pre-partition counts, not final training-set sizes.
Geographically Grounded Associations
Recovering a common spatial reference. Source meshes use different coordinate systems, units, local origins, and axis conventions. We retain source metadata and recover the transform from local geometry to the geographic image frame before extracting instance observations. Coordinate projection, axis order, local offsets, and units are resolved in this transformation chain. The image affine mapping is updated after cropping, padding, and resizing so that pixel spacing continues to represent the correct physical extent. Meshes and partial points are placed in the same upright building coordinate frame for training. Source-to-geographic transforms remain attached to the association rather than being inferred from normalized geometry.
Retrieving overhead evidence. A building’s geographic extent defines a context window in an orthophoto or remote-sensing mosaic. We retain the image-to-world transformation and available source metadata. Municipal orthophotos and mixed aerial/satellite basemaps are both used; consequently, we describe the collection as overhead remote-sensing imagery rather than assuming a uniform satellite sensor. The context preserves neighboring structures, while an instance mask focuses attention on the target.
Instance masks and paired conditions. We project the source building geometry into the geographic window to initialize an instance mask. RGB context, mask, and cutout are stored as related observations of the same target. Partial point clouds use available measurements or source-derived simulations according to their source pool. Coarse geometric conditions include bounding volumes and footprint-derived prisms. A prior that uses target-derived height is distinguished from one derived entirely from external observations; it is not treated as free information in image-only reconstruction. The projected instance mask is not supplied as an additional condition in the single-image setting.
For quantitative evaluation, each target building is localized using a 2D bounding box obtained by projecting its reference geometry into the georeferenced image. The same bounding-box crop is used for all compared methods. In the single-image setting, background pixels inside the crop are retained. The projected instance mask and its bounding box are distinct annotations: storing a mask does not imply that background removal is applied in every input setting. Our image-based evaluation therefore assumes ground-truth-derived instance localization and measures reconstruction given an instance crop, rather than end-to-end building detection and reconstruction. Although this is sufficient for single-building generation, we are also exploring, at this stage, the use of agents to automate the entire city-scale generation pipeline.
Geometric preparation. We remove invalid or empty geometric assets, standardize orientation, and encode the valid surfaces into sparse-structure and shape representations. Base-plane handling follows the source’s conventions and the open-bottom training target; it is not a general watertightness repair. Geometry is normalized consistently with its associated point observations.
Quality Screening and Refinement
Native image information. Quality screening checks native resolution, foreground coverage, exposure, texture, and orientation. Resolution tiers use the source image before resampling; enlarging a crop does not add evidence or improve its tier. Context and cutout variants share the same image window, padding, and display scale, with only the instance mask differing.
Coordinate, unit, and scale calibration. Systematic scale disagreement is first diagnosed in the coordinate conversion, rather than absorbed by a per-building image fit. Incorrect map projection, confusion between meters and feet, or a stale pixel-to-world scale after crop resizing can affect every target from a source. Our Boston repair restores the local offset and the source survey-foot convention. The inspected Cambridge export additionally requires a source-specific scale factor; it is not a universal correction for Cambridge data. These corrections belong to the shared source-to-world transform and are checked against geographic context before local alignment.
Different boundaries and height parallax. A visible roof boundary, an eave outline, and a ground footprint need not coincide. Projecting a source mesh into an image can also disagree with visible edges because an oblique acquisition displaces elevated surfaces relative to the ground, with different roof levels shifting differently. Occlusion and differences in acquisition date introduce further discrepancies. A single isotropic scale cannot generally resolve these effects. We therefore distinguish coordinate calibration from correspondence refinement, retain the meaning of each contour, and flag ambiguous roof–footprint matches instead of treating every residual as a scale error.
Shape-preserving boundary refinement. After geographic calibration, local refinement addresses residual contour displacement where the projected and visible boundaries are compatible. Our refinement proposal searches a bounded translation and isotropic scale around the registered contour. For contour point and its center , the candidate transform is . The objective combines orientation-compatible image-edge distance with penalties on translation and scale , preserving the source polygon rather than reclassifying foreground pixels. The transformation preserves polygon components, vertex ordering, and corners, avoiding the erosion of roof structure that unconstrained image segmentation can introduce. We retain a proposal only when it improves the edge objective and remains inside the image; otherwise the registered prior is kept. Successive refinement passes record their cumulative scale, so a bound relative to an intermediate contour is not confused with a bound relative to the source geometry. Figure S2 shows examples.

Figure S2. Preserving building structure during image alignment. Six cities, with initial contours and boxes (red), registered candidates (green), candidate cutouts, and source-mesh top views. The shared image window preserves visual scale across the first three columns. Source meshes use a strict top-down view to check structure, not an estimated satellite camera. Local refinement preserves polygon structure; rows can retain an already registered input when no additional update is accepted. These curation examples are separate from the observations used to train the evaluated checkpoints; they do not provide independent mask annotations.
Refinement preserves vector-contour structure and retains the registered prior when a proposed update fails the acceptance criteria. No independent pixel annotations are available for this refinement, so the examples do not establish a mask-IoU improvement. Difficult cases involving occlusion, multiple roof levels, building groups, or an incorrect source contour remain subject to review.
Splits and Representative Cases
The evaluation split contains 2,000 unique building targets in two groups. The unseen-city group contains 250 buildings from each of Greater Geelong, Philadelphia, Perth, and Toronto, for 1,000 targets. These four cities are excluded from training across images, point clouds, geometry, and derived observations.
The unseen-region group contains 100 buildings from each of Calgary, Wellington, Tokyo, Montreal, Hong Kong, Cape Town, Edmonton, Mississauga, New York, and Melbourne, for another 1,000 targets. Contiguous geographic regions are reserved before selecting the test buildings. All buildings in these regions are excluded from training, not only the selected test targets. Spatial buffers cover at least the image-crop extent to prevent overlap between training and test imagery.
The four held-out cities account for 10,753 image–mesh pairs in Table S1, leaving 1,000,443 pairs before region and buffer exclusions. The resulting training set contains approximately one million image–mesh pairs. Training uses eligible targets outside the excluded cities, regions, and buffers. All image, point-cloud, and prior variants of a building remain in the same partition. The training IDs, test IDs for both groups, and geographic exclusion lists are frozen. The 18-city counts in Table S1 precede these exclusions. The query-resolution, surface-stage, and paired degradation experiments use the full 2,000-building test set.
Figure S1 illustrates associations across 15 cities, which are not restricted to the training subset. It pairs geographic context and source geometry with individual building observations. Figure S2 and Figure S3 show alignment examples and textured source meshes, respectively. These figures illustrate data construction and source assets, rather than model predictions or a separate test set.

Figure S1. BuildingWorld-Image across 15 cities. Each city panel pairs a geographic image window with source building geometry and shows a selected instance crop with its corresponding mesh. The examples illustrate geographic and architectural diversity; they are not restricted to the training subset. Road traces under rendered meshes provide schematic visual context and are not surveyed road annotations. Building geometry is unchanged.
Image-Based Evaluation and Rendering
The main-text image-generation table reports two input settings with different image preprocessing protocols. The single-image comparison uses real remote-sensing instance crops with the background retained. It evaluates building-fine-tuned TRELLIS, TRELLIS.2, SAM 3D Objects, and Any2Building using the same input image and crop. The prior-assisted setting instead uses background-removed instance crops from orthophotos, reflecting OSM-assisted reconstruction scenarios with available building footprints. Training in this setting uses the corresponding background-removed images. The comparison evaluates SpaceControl, Hunyuan3D-Omni, SatSkylines, and Any2Building using the same image crop and coarse geometric prior. Within each setting, all methods share the input preprocessing, reference geometry, and evaluation frame.
Texture generation. All compared methods follow a reference-image-assisted texturing workflow. For each predicted untextured mesh, appearance-reference images are synthesized and used to condition the corresponding texture-generation module. Each baseline retains its own texture generator. The synthesized appearance references used for conditioning are distinct from the ground-truth renderings used for evaluation. Accordingly, appearance scores evaluate the complete geometry-and-texturing pipelines, rather than isolating the contribution of the geometry generator.
Appearance evaluation. PSNR and SSIM are evaluated on the subset with textured reference geometry. Following the four-view camera configuration of SatSkylines, we render each predicted textured mesh and its reference from yaw angles , with a fixed pitch of , field of view of , and camera radius of 2 in the normalized rendering frame. Predictions and references use identical camera poses, rendering resolution, background, and lighting.
We compute PSNR and SSIM between corresponding predicted and reference renderings, average the four view scores for each building, and then average across buildings. RGB values are normalized to . These metrics compare rendered textured meshes rather than renderings against the input remote-sensing image, and therefore reflect both geometric and appearance differences.
Implementation, Evaluation, and Reconstruction
Architecture, Training, and Inference
The point encoder uses four EdgeConv stages with 16-neighbor aggregation and farthest-point reductions to 2,048 and 1,024 centers. It returns 1,024 spatially indexed tokens of width 1,024. The frozen DINOv3 encoder processes images at . The spatial-query lattice contains locations. Query features are projected into a 30-block diffusion transformer of hidden width 1,536. Spatial prompts enter blocks 0, 10, 20, and 29 (zero-based indices), while query-enhanced condition tokens supply cross-attention throughout the transformer.
The shared training configuration samples point-only (P), image-only (I), point–image (P+I), and image-plus-prior (I+R) modes with equal probability. Point degradation precedes resampling to the encoder input size; repeated input rows do not constitute additional observations. The query occupancy loss has weight 0.1. Voxel-level geometric supervision uses the frozen sparse-structure decoder with a BCE/Dice auxiliary loss weighted by 0.02. Query-enhanced tokens and spatial prompts are dropped together with probability 0.1 for classifier-free guidance.
At inference, the surface generator operates at the predicted occupied voxel coordinates, without additional image or point-cloud features. It is therefore conditioned on the predicted structure, even though direct observation conditioning is removed.
Baseline adaptation. We initialize the learned 3D generation baselines from their released pretrained checkpoints and fine-tune them on the training split corresponding to each experiment. Each independently trained generative stage receives 200,000 optimizer updates, using the task-specific input preprocessing described in Appendix § 6.§ 5. For Hunyuan3D-Omni, we fine-tune the shape diffusion model and unified control encoder jointly as a single shape-generation stage. The update budget refers to optimizer steps rather than inference sampling steps.
Baseline availability. BuilDiff and Axolotl3D are discussed as related work but are excluded from the quantitative comparisons because we could not obtain runnable official implementations as of November 15, 2026. BuilDiff’s official repository contained only a README and license, while the Axolotl3D project page did not provide a public implementation or checkpoint download.
Geometric Metrics and Output Representations
The Building-PCC comparison and encoder ablation follow the metric definitions and reporting scale of BuildAny-Point [13]. Its supplementary material defines CD as the sum of the two directional mean nearest-neighbor L1 (Manhattan) distances, using 16,384 points per cloud. The geometric distance definitions and scaling detailed below apply to BuildingWorld-Image.
In the completion comparisons, PC denotes completed point sets before surface generation and low-poly reconstruction, whereas LP denotes final low-poly meshes. Mesh-based geometric metrics are computed from surface-sampled points.
Reference meshes are sampled uniformly by surface area with 16,384 points. A common evaluation frame uses the reference centroid and maximum radius for every method. For predicted points and reference points , define . We compute
and
Precision is the fraction of predicted points within of the reference, and recall is defined in the reverse direction. The surface F-score is
with when . Unless stated otherwise, reported surface F-scores use .
EMD uses exact bijective matching on 1,024 sampled points and averages two sampling seeds. CD and EMD in this evaluation are multiplied by . For point outputs, CD and F-score use the native prediction point count. Uniformity and DCD use 16,384 points, with replacement only when necessary. Uniformity is the normalized variance of neighbor counts within radius 0.05 at 100 query points, averaged over three seeds. DCD uses and count exponent 1.
Voxel F1 in Appendix § 8.§ 2 measures overlap between predicted and reference surface-support voxels. These voxels represent building surfaces, not filled interiors. Voxel F1 and surface F-score measure different quantities and are not numerically interchangeable.
Low-poly complexity. For the Building-PCC low-poly results in Table 1, we compare predicted and reference vertex and face counts. For building and count type , let and denote the corresponding counts, and define . Over evaluated buildings, we report
Each table entry reports with in parentheses. The error is averaged over individual buildings rather than computed from the aggregate ratio, preventing overestimation and underestimation across buildings from cancelling out.
Ablation Protocols
Evaluation scope. All three studies in Table 4 use the full 2,000-building BuildingWorld-Image test set. Geometric metrics are evaluated on the final low-poly meshes produced by the reconstruction procedure specified for each study, using the shared evaluation frame and metric definitions in Appendix § 7.§ 2. F-score uses a normalized distance threshold of 0.02, and is reported with a scale factor of . Component variants are compared within each input setting in panel (a); the paired contribution of imagery is evaluated separately in panel (b). In the surface-generation study, both the structure-only and surface-stage variants are converted to low-poly meshes using the indicated backend before evaluation.
Spatial-query variants. The condition concat baseline replaces the spatial-query conditioning design with direct concatenation of the available input tokens. Point and image tokens are included according to the input setting, and encoded prior tokens are also concatenated in the I+R setting. The resulting sequence directly conditions the DiT, without query read-back or spatial-prompt injection. This baseline therefore compares the complete spatial-query conditioning design with direct token conditioning while retaining access to the available input modalities and priors.
The remaining variants remove one component at a time. The w/o occupancy supervision variant omits the auxiliary query occupancy loss while retaining the query representation and the remaining generation objectives. The w/o query read-back variant uses the base condition tokens directly, i.e., , while retaining query occupancy supervision and spatial prompts. The w/o spatial prompts variant removes the aligned residual injection into the DiT while retaining query-enhanced condition tokens and occupancy supervision. Each variant is retrained using the same training protocol; the components are not simply disabled in the trained full model at inference.
Joint and single-task training. We train one specialist for each of , , , and , using the corresponding architecture and inference inputs. Each specialist and the joint model receive the same number of optimizer updates. The joint model samples the four input modes with equal probability, whereas each specialist uses only its designated mode. The comparison therefore matches the total update count per model rather than the number of training exposures to each mode. At evaluation, the joint model and the corresponding specialist receive the same type of input; joint training does not introduce extra inference modalities into the P-only or I-only comparisons.
Paired point degradation. The image-contribution study uses the same trained checkpoint for P and P+I inference. The None setting uses the original 2,048-point partial observation without additional degradation; it does not denote a complete building point cloud. Random degradation retains 205 or 41 points from this observation. Spatial degradation adapts the ShapeNet-55 cropping scheme used by AdaPointR [39]. A randomly selected spatial region is first removed, after which the remaining observations are subsampled to 41 points. Thus, spatial removal precedes subsampling, and the 41-point input is not obtained by directly cropping away all but 41 points. Comparing this setting with random subsampling to the same count examines the role of spatial coverage under equally sparse observations.
For every comparison, P and P+I use identical retained points, normalization, and generation noise; P+I additionally receives the paired overhead image. Degradation is applied before resampling to the encoder input size, so repeated samples do not add geometric observations. The reported point counts refer to the retained observations rather than the length of the resampled encoder input.
Surface-stage variants and their point-conditioning comparison are described in Appendix § 7.§ 5; conversion and timing settings are specified in Appendix § 7.§ 4.
Low-Poly Conversion and Timing Protocol
Point-output baselines use a fixed marching-cubes adapter before low-poly conversion. The structure-only variant uses its predicted support mesh, and the complete pipeline uses the structure-conditioned surface generator. BuildAnyPoint supplies its native completion-to-mesh result, which is repeated as a reference in both backend groups of Table S2.
| Method | PolyFit Time (s) | PolyFit CD | PolyFit F-score | RoLoPM Time (s) | RoLoPM CD | RoLoPM F-score |
| FoldingNet [37] | 121.29 | 48.02 | 0.42 | 239.08 | 51.76 | 0.31 |
| PCN [40] | 116.92 | 55.86 | 0.33 | 219.13 | 56.53 | 0.38 |
| TopNet [27]‡ | 217.59 | 87.90 | 0.09 | 382.77 | 64.25 | 0.21 |
| GRNet [34] | 125.69 | 70.88 | 0.33 | 215.34 | 54.77 | 0.30 |
| SnowflakeNet [33] | 62.57 | 35.98 | 0.67 | 186.02 | 11.69 | 0.91 |
| PoinTr [38] | 73.33 | 42.55 | 0.58 | 204.71 | 12.68 | 0.88 |
| AnchorFormer [3] | 61.12 | 51.11 | 0.62 | 167.30 | 13.87 | 0.85 |
| AdaPoinTr [39] | 60.91 | 40.14 | 0.62 | 161.57 | 11.05 | 0.92 |
| BuildAnyPoint [13]† | 38.39 | 12.88 | 0.92 | 38.39 | 12.88 | 0.92 |
| Ours (Structure only) | 7.94 | 10.10 | 0.9387 | 91.20 | 9.83 | 0.9588 |
| Ours (Structure + Surface) | 13.23 | 9.06 | 0.9472 | 97.57 | 8.65 | 0.9719 |
Table S2. Low-poly reconstruction. Arithmetic means on the full 2,000-building BuildingWorld-Image test set, with the TopNet CD exception specified below. CD denotes CD . F-scores are reported as fractions at a normalized distance threshold of 0.02, to four decimal places for Any2Building and two for the baselines. Time sums recorded inference and conversion times.
RoLoPM uses a screen-size setting of 100 and an intermediate budget of 128k triangles. PolyFit uses 20k surface samples with fixed plane-detection and optimization settings, including one common coarse-plane retry upon failure. For the PolyFit adapter, open-bottom surface-stage outputs receive 4k bottom-support samples derived from their predicted minimum height. These representation-specific adapters remain fixed within the comparison.
Reported time sums cached GPU inference and measured CPU conversion using eight cores. It excludes initialization, I/O, cleanup/export, metric computation, and evaluation alignment. It is an aggregate of recorded components rather than a newly measured end-to-end wall-clock time. Results are reported in Appendix § 7.§ 5.
Surface Generation and Low-Poly Reconstruction
Table S2 extends the main-text surface-stage comparison to additional point-cloud completion baselines, using the conversion and timing protocol in Appendix § 7.§ 4. All methods are evaluated on the full 2,000-building BuildingWorld-Image test set, with shared conversion settings within each backend.
The surface study in Table 4(c) is conducted in the point-cloud-input setting and uses the same predicted building structures as inputs to the reconstruction paths. The Structure only variant converts these structures to low-poly meshes without the learned surface stage. The two + Surface variants insert surface generation before low-poly reconstruction. At inference, both surface generators are conditioned on the predicted occupied voxel coordinates, while the w/ point cond. variant additionally receives point-cloud conditioning. This additional conditioning branch is included when training the corresponding surface generator. The w/o point cond. variant uses only the predicted voxel coordinates at inference and is therefore not unconditional. PolyFit and RoLoPM are evaluated separately, following the fixed conversion protocols in Appendix § 7.§ 4.
Adding surface generation reduces PolyFit CD from 10.10 to 9.06 and increases F-score from 0.9387 to 0.9472. For RoLoPM, CD decreases from 9.83 to 8.65 and F-score increases from 0.9588 to 0.9719. The corresponding time increases are 5.29 s and 6.37 s. The complete pipeline achieves the lowest reported CD and highest F-score with each backend on this cohort.
Among methods using each conversion backend, the structure-only variant has the shortest recorded time. The native BuildAnyPoint result takes 38.39 s and is shown as a reference in both column groups. The reported times reflect the combined costs of inference, adapter construction, and low-poly conversion.
Additional point-cloud conditioning. Table 4(c) shows that additional point-cloud conditioning yields slightly worse reported geometry metrics with both backends. With PolyFit, CD increases from 9.06 to 9.21 and F-score decreases from 0.9472 to 0.9458; with RoLoPM, the corresponding changes are 8.65 to 8.72 and 0.9719 to 0.9706. Recorded runtime also increases from 13.23 to 16.12 seconds with PolyFit and from 97.57 to 99.64 seconds with RoLoPM. These observations show no reconstruction benefit from injecting the point-cloud condition again at the surface stage in this experiment. The reported time combines inference and low-poly conversion and follows the timing convention in Appendix § 7.§ 4; it does not isolate the execution time of the conditioning branch.
Qualitative surface comparison. In the surface-generation panel of the main-text component figure, the Stage-1 support is held fixed. The displayed Stage-2 outputs were selected from multiple generation seeds to reduce visible surface defects; the geometry itself was not filled or edited. These are illustrative outputs, not an estimate of average stochastic performance. The quantitative surface-stage results in Table S2 use the recorded evaluation outputs, not the seeds selected for display.
Additional Ablations and Analysis
This section examines point-cloud encoding, query resolution, and geometric priors, and provides qualitative component comparisons and textured visualizations. The encoder ablation uses Building-PCC; the query-resolution experiment uses the full BuildingWorld-Image test set. The paired point-degradation protocol is described in Appendix § 7.§ 3.
Point-Cloud Encoder Ablation
We compare different point-cloud conditioning encoders on Building-PCC while keeping the spatial-query module and generation pipeline fixed. Specifically, we replace EdgeConv with a pointwise MLP, a Point Transformer, or the TRELLIS.2 VAE encoder [32]. The point-based variants use the same sampled centers, whereas the VAE encoder extracts conditioning features from voxelized partial observations. All encoder outputs are adapted to the same token count and feature width. We additionally remove the additive positional embedding from the EdgeConv tokens to isolate its contribution while retaining local neighborhood aggregation. All variants follow the same training split, optimization budget, and evaluation protocol. Table S3 compares point-cloud completion and downstream low-poly reconstruction to assess the effects of local geometry encoding, voxel-based conditioning, and explicit point localization.
| Variant | PC F-score | PC CD | Uniformity | EMD | LP F-score | LP CD |
| TRELLIS.2 VAE encoder | 0.89 | 0.38 | 0.08 | 0.18 | 0.86 | 0.48 |
| Pointwise MLP | 0.91 | 0.33 | 0.07 | 0.15 | 0.88 | 0.44 |
| Point Transformer | 0.93 | 0.27 | 0.05 | 0.12 | 0.90 | 0.38 |
| EdgeConv w/o PE | 0.95 | 0.22 | 0.04 | 0.09 | 0.93 | 0.33 |
| EdgeConv (Ours) | 0.96 | 0.20 | 0.03 | 0.07 | 0.94 | 0.30 |
Table S3. Point-cloud encoder ablation on Building-PCC. Metrics follow the Building-PCC comparison in the main text (PC: point cloud; LP: low-poly reconstruction). F-scores are reported as fractions rounded to two decimal places. Each variant replaces the point-cloud conditioning encoder or removes its additive positional embedding (PE). The spatial-query module, generative backbone, output decoder, and downstream reconstruction stages remain unchanged.
EdgeConv with positional embeddings has the highest PC and LP F-scores and the lowest reported distance and Uniformity values among these variants. Removing its positional embedding reduces PC F-score from 0.96 to 0.95 and LP F-score from 0.94 to 0.93. This comparison isolates the additive center embedding while retaining neighborhood aggregation.
Spatial-Query Design and Resolution
To examine the spatial granularity of the query design, we compare and query lattices through controlled continuation from a common initialization, matching the number of additional optimizer steps. The coarser query field
Bold marks the best value in each column among methods using the corresponding conversion backend. The surface generator in our full pipeline uses only the predicted voxel coordinates as conditioning. †Both column groups show the same native BuildAnyPoint result as a reference. ‡PolyFit succeeds on approximately half of the TopNet outputs. Failed conversions receive zero F-score. CD averages only successful conversions, and the mean time includes all attempts. Timing exclusions and adapters are specified in Appendix § 7.§ 4.
is interpolated to the generator’s latent grid, while the output resolution remains in both variants. Query supervision and prior aggregation follow the query resolution, so this experiment evaluates the spatial-query configuration as a whole rather than query count alone.
Both variants use EMA weights, 12 sampling steps, guidance 3.0, and paired observations and noise on the full 2,000-building BuildingWorld-Image test set. Table S4 reports voxel F1 for point-only (P), image-only (I), joint point–image (P+I), and image-plus-prior (I+R) conditioning.
| Query grid | P | I | P+I | I+R |
| 0.7703 | 0.2213 | 0.7671 | 0.2681 | |
| 0.7723 | 0.2272 | 0.7676 | 0.2987 |
Table S4. Query-resolution comparison on the full 2,000-building BuildingWorld-Image test set. Voxel F1 measures surface-support overlap; higher is better.
The finer lattice yields slightly higher mean voxel F1 in all four modes, with the largest observed gain under image-plus-prior conditioning. For , the mean paired differences are +0.0020 (P), +0.0059 (I), +0.0004 (P+I), and +0.0306 (I+R). Differences are computed before rounding and may therefore differ slightly from subtraction of the displayed table entries.
Image-Conditioned Geometric Priors
Prior-assisted generation on Skylines-50K. Following the 500-asset evaluation setting of SatSkylines [18], we reproduce the compared methods on Skylines-50K and evaluate LoD0 and LoD1 priors separately. As shown in Table S5, Any2Building achieves the best results across all geometry and appearance metrics under both prior settings, demonstrating effective building reconstruction with geometric constraints of varying detail.
| Method | Geometry | Appearance | ||||
| IoU | CD | F-score | PSNR | LPIPS | CLIP | |
| (a) LoD0 priors | ||||||
| SpaceControl [7] | 0.6318 | 0.0402 | 0.5924 | 16.2318 | 0.2714 | 0.6253 |
| Hunyuan3D-Omni [28] | 0.8705 | 0.0268 | 0.7920 | 17.4162 | 0.2187 | 0.7468 |
| SatSkylines [18] | 0.8982 | 0.0224 | 0.8236 | 17.9248 | 0.2015 | 0.7601 |
| Ours | 0.9307 | 0.0172 | 0.8702 | 19.5238 | 0.1624 | 0.7889 |
| (b) LoD1 priors | ||||||
| SpaceControl [7] | 0.6612 | 0.0357 | 0.6218 | 16.8712 | 0.2481 | 0.6432 |
| Hunyuan3D-Omni [28] | 0.8894 | 0.0231 | 0.8107 | 18.0512 | 0.1986 | 0.7590 |
| SatSkylines [18] | 0.9165 | 0.0189 | 0.8421 | 18.6374 | 0.1813 | 0.7742 |
| Ours | 0.9452 | 0.0139 | 0.8906 | 20.4162 | 0.1402 | 0.8015 |
Table S5. Prior-assisted building generation with LoD0 and LoD1 priors. Results follow the 500-asset evaluation setting and are compared separately within each prior level. F-scores are reported as fractions. PSNR is reported in dB.
A rough prior constrains the feasible building volume while leaving the generator to infer roof structure and facade articulation. Figure S4 compares image-only generation with prior-assisted generation on selected examples. The examples show roof structure beyond the coarse prior, but do not quantify a mean improvement over a matched image-only baseline.
Figure S4. Geometric priors reduce image-to-3D ambiguity. Selected image-only and image-plus-prior reconstructions, each shown with its reference geometry. Priors help anchor building extent and massing while the generator recovers architectural detail. These examples illustrate the additional-input setting rather than image-only performance.
Prior quality and provenance matter: target-derived heights constitute extra supervision at inference, and incorrect footprints or heights can bias the result. We therefore distinguish the prior-conditioned setting from image-only reconstruction throughout.
Qualitative Interpretation of Components
Figure 7(b) and Figure S5 visualize query occupancy and decoded building support alongside the reference surface voxels. Figure 7(c) shows predictions from single-task and jointly trained models. These examples complement the quantitative component comparisons in Table 4(a). The surface-generation comparison in Figure 7(a) is discussed in Appendix B.5.
Figure S5. Spatial queries provide an interpretable geometric intermediate. From left to right: decoder support without queries, query anchors with occupancy support, occupied query cells, decoder support with queries, and reference support. The query lattice is , with occupied cells visualized at probability at least 0.5. Each row preserves a common view. Quantitative component comparisons are reported in Table 4(a).
Downstream Textured Visualization
Figure S6 illustrates downstream textured visualization, pairing input imagery and untextured building meshes with reference appearances and final textured outputs. These qualitative examples complement the reconstruction results

Figure S6. Downstream textured building visualization. Each group shows the input image, untextured mesh (White Mesh), reference appearance, and final textured output. The examples illustrate the appearance of the textured buildings alongside their geometric structure.
References
- [1]Xingyu Chen, Fu-Jen Chu, Pierre Gleize, Kevin J Liang, Alexander Sax, Hao Tang, Weiyao Wang, Michelle Guo, Thibaut Hardin, Xiang Li, Aohan Lin, Jia-Wei Liu, Ziqi Ma, Anushka Sagar, Bowen Song, Xiaodong Wang, Jianing Yang, Bowen Zhang, Piotr Dollár, Georgia Gkioxari, Matt Feiszli, and Jitendra Malik. SAM 3D: 3Dfy anything in images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7220–7232, 2026. 2, 3, 7arxiv.org/abs/2511.16624
- [2]Yingshu Chen, Huajian Huang, Tuan-Anh Vu, Ka Chun Shum, and Sai-Kit Yeung. StyleCity: Large-scale 3D urban scenes stylization. In European Conference on Computer Vision (ECCV), pages 395–413, 2024. 2arxiv.org/abs/2404.10681
- [3]Zhikai Chen, Fuchen Long, Zhaofan Qiu, Ting Yao, Wengang Zhou, Jiebo Luo, and Tao Mei. AnchorFormer: Point cloud completion from discriminative nodes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13581–13590, 2023. 3, 6, 7, 21DOI
- [4]Zhen Chen, Zherong Pan, Kui Wu, Etienne Vouga, and Xifeng Gao. Robust low-poly meshing for general 3D models. ACM Transactions on Graphics, 42(4):119:1–119:20, 2023. 3, 5, 8DOI
- [5]Zhaiyu Chen, Yuqing Wang, and Xiao Xiang Zhu. Unified primitive proxies for structured shape completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7457–7467, 2026. 3, 6, 7arxiv.org/abs/2601.00759
- [6]Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G. Schwing, and Liang-Yan Gui. SDFusion: Multimodal 3D shape completion, reconstruction, and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4456–4465, 2023. 2, 3arxiv.org/abs/2212.04493
- [7]Elisabetta Fedele, Francis Engelmann, Ian Huang, Or Litany, Marc Pollefeys, and Leonidas Guibas. SpaceControl: Introducing test-time spatial control to 3D generative modeling. In International Conference on Learning Representations (ICLR), 2026. 3, 7, 22arxiv.org/abs/2512.05343
- [8]Wenxiao Gao, Ravi Peters, and Jantien Stoter. Building-PCC: Building point cloud completion benchmarks. ISPRS Annals of Photogrammetry, Remote Sensing and Spatial Information Sciences, X-4/W5-2024:179–186, 2024. 6arxiv.org/abs/2404.15644
- [9]Hongye Hou, Yuehao Gao, Zhan Liu, and Yang Yang. Dig into detailed structures: Key context encoding and semantic-based decoding for point cloud completion. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6686–6695, 2024. 3
- [10]Hongye Hou, Yang Yang, Zhan Liu, and Fangyu Du. Benefit from reference: Retrieval-augmented cross-modal point cloud completion. Applied Soft Computing, 201:115626, 2026. 3arxiv.org/abs/2507.14485
- [11]Anita Hu and Maria Shugrina. Axolotl3D: A unified framework for faithful 3D shape completion. In European Conference on Computer Vision (ECCV), pages 1–20, 2026. 2, 3arxiv.org/abs/2607.20660
- [12]Tongyan Hua, Lutao Jiang, Ying-Cong Chen, and Wufan Zhao. Sat2City: 3D city generation from a single satellite image with cascaded latent diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 27978–27988, 2025. 3arxiv.org/abs/2507.04403
- [13]Tongyan Hua, Haoran Gong, Yuan Liu, Di Wang, Ying-Cong Chen, and Wufan Zhao. BuildAnyPoint: 3D building structured abstraction from diverse point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17099–17109, 2026. 2, 3, 6, 7, 18, 21arxiv.org/abs/2602.23645
- [14]Tongyan Hua, Dongli Wu, Jinjing Zhu, Yinrui Ren, Zhongcheng Hong, Ying-Cong Chen, Hui Xiong, and Wufan Zhao. Sat2City v2: Native 3d city asset generation from a single satellite image. arXiv preprint arXiv:2606.24138, 2026. 3arxiv.org/abs/2606.24138
- [15]Jin Huang, Jantien Stoter, Ravi Peters, and Liangliang Nan. City3D: Large-scale building reconstruction from airborne LiDAR point clouds. Remote Sensing, 14(9):2254, 2022. 2, 3, 6DOI
- [16]Qirui Huang, Runze Zhang, Kangjun Liu, Minglun Gong, Hao Zhang, and Hui Huang. ArcPro: Architectural programs for structured 3D abstraction of sparse points. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6563–6572, 2025. 2, 3arxiv.org/abs/2503.02745
- [17]Shangfeng Huang, Ruisheng Wang, and Xin Wang. Building-World: A structured 3D building dataset for urban foundation models. Proceedings of the AAAI Conference on Artificial Intelligence, 40(7):5085–5094, 2026. 2arxiv.org/abs/2511.06337
- [18]Zhangyu Jin and Andrew Feng. SAT-SKYLINES: 3D building generation from satellite imagery and coarse geometric priors. In International Conference on 3D Vision (3DV), pages 1608–1618, 2026. 2, 3, 6, 7, 21, 22
- [19]Chieh Hubert Lin, Hsin-Ying Lee, Willi Menapace, Menglei Chai, Aliaksandr Siarohin, Ming-Hsuan Yang, and Sergey Tulyakov. InfinCity: Infinite-scale city synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22808–22818, 2023. 3DOI
- [20]Moyun Liu, Ziheng Yang, Bing Chen, Youping Chen, Jingming Xie, Lei Yao, Lap-Pui Chau, Jiawei Du, and Joey Tianyi Zhou. MAENet: Boost image-guided point cloud completion more accurate and even. Information Fusion, 122:103179, 2025. 3, 7DOI
- [21]Xinpu Liu, Baolin Hou, Hanyun Wang, Ke Xu, Jianwei Wan, and Yulan Guo. DulNet: Dual-modality feature interaction for point cloud completion. IEEE Transactions on Multimedia, 27:8525–8540, 2025. 3, 7arxiv.org/abs/2407.07374
- [22]Yujia Liu, Anton Obukhov, Jan Dirk Wegner, and Konrad Schindler. Point2Building: Reconstructing buildings from airborne LiDAR point clouds. ISPRS Journal of Photogrammetry and Remote Sensing, 215:351–368, 2024. 2, 3, 6
- [23]Liangliang Nan and Peter Wonka. PolyFit: Polygonal surface reconstruction from point clouds. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2353–2361, 2017. 3, 8DOI
- [24]Ming Qian, Zimin Xia, Changkunk Liu, Shuailiei Ma, Wen Wang, Zeran Ke, Bin Tan, Hang Zhang, and Gui-Song Xia. Sat3DGen: Comprehensive street-level 3D scene generation from single satellite image. In International Conference on Learning Representations (ICLR), 2026. 3
- [25]Xuanchi Ren, Jiahui Huang, Xiaohui Zeng, Ken Museth, Sanja Fidler, and Francis Williams. XCube: Large-scale 3D generative modeling using sparse voxel hierarchies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4209–4219, 2024. 2arxiv.org/abs/2312.03806
- [26]Oriane Siméoni, Huy V. Vo, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Sfrzaniec, Seung Eun Yi, Michael Ramamonjisoa, Francisco Massa, Daniel Haziza, Luca Wehrstedt, Jianyuan Wang, Timothée Darcet, Théo Moutakanni, Leonel Sentana, Claire Roberts, Andrea Vedaldi, Jamie Tolan, John Brandt, Camille Couprie, Julien Mairal, Hervé Jégou, Patrick Labatut, and Piotr Bojanowski. DINOv3. Transactions on Machine Learning Research, 2026. 3arxiv.org/abs/2508.10104
- [27]Lyne P. Tchapmi, Vineet Kosaraju, Hamid Rezatofighi, Ian Reid, and Silvio Savarese. TopNet: Structural point cloud decoder. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 383–392, 2019. 3, 6, 21DOI
- [28]Team Hunyuan3D, Bowen Zhang, Chunchao Guo, Haolin Liu, Hongyu Yan, Huiwen Shi, Jingwei Huang, Junlin Yu, Kunhong Li, Linus, Penghao Wang, Qiqiang Lin, Sicong Liu, Xianghui Yang, Yixuan Tang, Yunfei Zhao, Zeqiang Lai, Zihao Liang, and Zibo Zhao. Hunyuan3D-Omni: A unified framework for controllable generation of 3D assets. arXiv preprint arXiv:2509.21245, 2025. 2, 3, 7, 22
- [29]Xinliang Wang, Yijie Kang, Zhenyu Wu, and Yifeng Shi. Sat2RealCity: Geometry-aware and appearance-controllable 3D urban generation from satellite imagery. arXiv preprint arXiv:2511.1470, 2025. 3arxiv.org/abs/2511.11470
- [30]Yao Wei, George Vosselman, and Michael Ying Yang. BuildDiff: 3D building shape generation using single-image conditional point cloud diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 2910–2919, 2023. 2, 3arxiv.org/abs/2309.00158
- [31]Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3D latents for scalable and versatile 3D generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21469–21480, 2025. 2, 3, 7arxiv.org/abs/2412.01506
- [32]Jianfeng Xiang, Yiqian Xue, Sicheng Xu, Ruicheng Wang, Zelong Lv, Yu Deng, Hongyuan Zhu, Yue Dong, Hao Zhao, Nicholas Jing Yuan, and Jiaolong Yang. Native and compact structured latents for 3D generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14419–14429, 2026. 2, 3, 5, 7, 20arxiv.org/abs/2512.14692
- [33]Peng Xiang, Xin Wen, Yu-Shen Liu, Yan-Pei Cao, Pengfei Wan, Wen Zhang, and Zhizhong Han. SnowflakeNet: Point cloud completion by snowflake point deconvolution with skip-transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5499–5509, 2021. 3, 6, 7, 21arxiv.org/abs/2108.04444
- [34]Haozhe Xie, Hongxun Yao, Shangchen Zhou, Jiageng Mao, Shenping Zhang, and Wenxiu Sun. GRNet: Gridding residual network for dense point cloud completion. In Computer Vision – ECCV 2020, pages 365–381, 2020. 3, 6, 21arxiv.org/abs/2006.03761
- [35]Haozhe Xie, Zhaoxi Chen, Fangzhou Hong, and Ziwei Liu. CityDreamer: Compositional generative model of unbounded 3D cities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9666–9675, 2024. 3arxiv.org/abs/2309.00610
- [36]Hang Xu, Chen Long, Wenxiao Zhang, Yuan Liu, Zhen Cao, Zhen Dong, and Bisheng Yang. Explicitly guided information interaction network for cross-modal point cloud completion. In European Conference on Computer Vision (ECCV), pages 414–432, 2024. 3, 7arxiv.org/abs/2407.02887
- [37]Yaoqing Yang, Chen Feng, Yiru Shen, and Dong Tian. FoldingNet: Point cloud auto-encoder via deep grid deformation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 206–215, 2018. 3, 6, 21arxiv.org/abs/1712.07262
- [38]Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. PoinTr: Diverse point cloud completion with geometry-aware transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12498–12507, 2021. 3, 6, 7, 21arxiv.org/abs/2108.08839
- [39]Xumin Yu, Yongming Rao, Ziyi Wang, Jiwen Lu, and Jie Zhou. AdaPoinTr: Diverse point cloud completion with adaptive geometry-aware transformers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(12):14114–14130, 2023. 3, 6, 7, 19, 21arxiv.org/abs/2301.04545
- [40]Wentao Yuan, Tejas Khot, David Held, Christoph Mertz, and Martial Hebert. PCN: Point completion network. In 2018 International Conference on 3D Vision (3DV), pages 728–737, 2018. 3, 6, 21arxiv.org/abs/1808.00671
- [41]Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3836–3847, 2023. 5arxiv.org/abs/2302.05543
- [42]Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. CLAY: A controllable large-scale generative model for creating high-quality 3D assets. ACM Transactions on Graphics (TOG), 43(4):120:1–120:20, 2024. 2, 3arxiv.org/abs/2406.13897
- [43]Qilin Zhang, Jinyu Zhu, Olaf Wysocki, Benjamin Busam, and Boris Jutzi. GS4City: Hierarchical semantic gaussian splatting via city-model priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 68–77, 2026. 2
- [44]Juelin Zhu, Shuaibeng Peng, Long Wang, Han Tan, Yu Liu, Maojun Zhang, and Shen Yan. LoD-Loc v2: Aerial visual localization over low level-of-detail city models using explicit satellite alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 26610–26621, 2025. 2DOI




