acceptodds
Withdrawn from ICLR 2027

UrBLEND: Building Model Synthesis From Incomplete Point Cloud With Gnervative Prios

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Building reconstruction often relies on point clouds from LiDAR or visual reconstruction frontends. Noise, occlusion, and uneven sampling make these observations incomplete and uncertain. Recovering complete geometry from such sparse evidence is underconstrained, motivating pretrained shape priors. To support learning and evaluation in this setting, we introduce CityEXP, a scalable data-construction pipeline and benchmark protocol. Agent-coordinated procedural city generation preserves building identities and detailed geometry; high-fidelity scene rendering and visual reconstruction produce incomplete point observations paired with these targets. Our baseline, UrBlend, adapts Hunyuan3D-Omni's pretrained shape prior to building observations. Prepared surfaces supply latent training targets across compatible geometry sources. A pretrained surface decoder produces meshes; separately trained structural decoders produce wireframes after task adaptation. On CityEXP, the baseline achieves canonical-shape of and Chamfer distance of . CityEXP and UrBlend establish a benchmark and unified baseline for prior-guided building reconstruction.

Primary area:
datasets and benchmarks
Keywords:
3D Generative Models, Building Geometry Synthesis, Sparse Geometric Observations, Shape Completion, Point Cloud Completion, Generative Shape Priors, Implicit 3D Representations, Signed Distance Fields

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_knJdOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper builds a synthetic benchmark called CityEXP and adapts Hunyuan3D-Omni into what they call, UrBlend, a baseline for reconstructing complete 3D buildings from incomplete point clouds derived from rendered images.

Critical strengths

straightforward and easy to understand what was done.

Critical weaknesses

It is a dataset and benchmark paper, but the main benchmark lacks meaningful baseline comparisons.

Table 2 reports only UrBlend on CityEXP—there is no native Hunyuan3D-Omni model adapted under the same protocol, nor an independent reconstruction method. This makes it hard to tell what the proposed changes actually contribute.

The BuildingWorld3D ablation provides limited reassurance: the gain over the stronger native baseline is just 0.09 percentage points in F.02 at validation-selected checkpoints, with a confidence interval crossing zero; even the final-update gain is only 0.29 points. Or is there any supplementary visualisation to support your superiority?

The BuildingGPT comparison addresses a separate wireframe task, not reconstruction on CityEXP.

Overall, the experiments do not convincingly establish the benefit of the adaptation or how existing methods perform on the paper’s central benchmark.

Questions for authors

The main benchmark appears to be dense surface completion from incomplete point clouds rather than LoD reconstruction. Since the main UrBlend branch outputs a dense surface mesh, why is it not compared against sota point-cloud completion or generative 3D reconstruction baselines under the same protocol?

Without an explicitly geographic output representation, such as a compact LoD/polygonal building mesh, it is unclear what makes this fundamentally different from a category-specific 3D completion benchmark beyond the use of building data.

Reviewer_t1AROpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces CityEXP, a procedural urban dataset pairing image-derived point clouds with detailed building geometry. It also presents UrBlend, which adapts Hunyuan3D-Omni using multiscale point cloud features for mesh reconstruction, with separately trained structural decoders for vertex and edge prediction.

Critical strengths
  1. CityEXP preserves building identities and reference geometry, enabling paired supervision even where observations are incomplete.

  2. The Urblend framework connects surface reconstruction and structural decoding through a shared latent representation.

  3. The evaluation examines both missing surface recovery and preservation of observed geometry, providing useful diagnostics beyond overall reconstruction scores.

Critical weaknesses
  • The point cloud conditioning seems too strong for the task. The authors describe reconstruction “from sparse, incomplete point observations” in line 58. However, the inputs in Figures 9, 11, and 13 appear dense and relatively complete. Figures 8 and 12 show some missing observations, but these inputs still appear more complete than those shown in BuildAnyPoint. This makes the emphasis on incomplete observations less convincing. The supplementary material shows more incomplete ALS/AHN inputs, but I could not find corresponding mesh or wireframe reconstruction results.

  • The dataset’s geometry diversity is insufficiently demonstrated. Although city sample assets are mentioned, their selection and diversity remain unclear. Most examples show box-shaped buildings or simple roofs. In the supplementary material, section “03 / THE CITY COLLECTION,” many buildings appear to share similar extruded footprints, uniform heights, and flat roofs. There is no architectural diversity comparison with UrbanScene3D, MatrixCity, or the datasets used by BuildingGPT and BuildAnyPoint. I would like to see more complex buildings in the visualization comparison section. The author mentioned 3DBAG dataset, then let's use it as an example. Can the author show building generation results on complicated churches like Oude Kerk Amsterdam, or domed structures like Renaissance Koepelkerk?

  • The benchmark comparisons are limited.

    • In Table 3, if I understand correctly, if Utonia encoder is removed, then it is pure Hunyuan3D-Omni right? Then the difference between Hunyuan3D-Omni and UrBlend is just less than 1% in F.02? That makes me question whether it is necessary to have the Utonia encoder. And why are other dense point cloud to mesh methods, such as MeshAnything V2, not included?
    • For wireframe reconstruction, why do Tables 10 and 11 evaluate only MunichWF rather than also CityEXP? BuildingGPT’s reported performance is substantially below its original paper. Maybe this gap needs a clearer explanation. Also, why BuildAnyPoint is not presented in this table?
  • The mesh generation quality is questionable. Figure 11 and the supplementary “Dataset 01 Mesh” examples show UrBlend's mesh generation results. From my point of view, it is too blurred and overly smoothed. As an informal check, I used a screenshot of an occluded building RGB image in your supplementary material as input to TRELLIS.2. Its generated mesh appeared much sharper and more detailed, using only image conditioning. Although this is not a controlled comparison, it raises questions about why the proposed method loses so much detail despite explicit point cloud conditioning. Also, as for the figure 11, I need to see other methods' results like Hunyuan3D-Omni or others, not just your results and GT.

  • The meaning and practical benefit of LoD need clarification. The author showed LoD generation results in the supplementary “Dataset 02 LoD”. But why the mesh generation results remain densely triangulated? In my understanding, LoD means that lower detail assets typically reduce polygon counts and rendering costs. But the predictions shown by the author are still highly dense, making me very confused about the meaning of LoD in this paper.

Questions for authors
  1. How does performance change as point clouds become more incomplete or sparse? Can you provide mesh and wireframe results for the ALS/AHN inputs shown in the supplementary material?
  2. How diverse are the building assets and generated geometries compared with existing datasets? Can you provide results on more complex buildings, including intricate roofs, churches, or domed structures?
  3. Given the small F.02 improvement over the stronger native Hunyuan3D-Omni baseline, what is the necessity of adding Utonia? How does the method compare with other point cloud to mesh approaches, such as MeshAnything V2?
  4. What explains the substantial difference between your BuildingGPT results and those reported in its original paper? Can you provide matched qualitative comparisons and explain why wireframe reconstruction is not also evaluated on CityEXP?
  5. What causes the substantial loss of facade and rooftop details in the predicted meshes? Can you include qualitative comparisons with Hunyuan3D-Omni to clarify the proposed method’s contribution?
  6. What definition of LoD is used, and how do the densely triangulated predictions support compact building representations? Can you report their vertex and face counts?
  7. Can you provide side-by-side visualization comparisons showing the input point-cloud condition, ground truth, UrBlend, and competing methods on the same examples? Specifically, please include Hunyuan3D-Omni and MeshAnything V2 for mesh reconstruction, and BuildingGPT and BuildAnyPoint for edge/wireframe reconstruction.
  8. Your title "UrBLEND: Building Model Synthesis From Incomplete Point Cloud With Gnervative Prios", I think it should be "Generative Priors".
Reviewer_ZN4uOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper first presents an agent-assisted procedural city generator, CityEXP, for automating urban building synthesis and creating benchmark data for evaluating point-to-building approaches. It also proposes UrBlend, which extends Hunyuan3D-Omni to the point-cloud-to-building setting, and conducts experiments mainly on BuildingWorld3D to demonstrate the capability of UrBlend. However, it remains unclear how significantly CityEXP contributes to incomplete-point-cloud-to-building synthesis, while the superiority and suitability of UrBlend for this task also require further and deep investigation.

Critical strengths

The main strength is the agent-assisted, Unreal-based procedural city generator, which enables the creation of potentially unlimited city scenes for various purposes. However, the data scale demonstrated in the paper is still limited, and its benefits for training and benchmarking are not sufficiently explored to constitute a strong contribution.

Critical weaknesses

Technical issues

  1. Despite the rich data elements provided by CityEXP compared with existing datasets such as MatrixCity and UrbanScene3D, it is critical to show that representative methods trained on CityEXP can generalize well to other datasets, given its claimed high quality and data volume. However, the work does not demonstrate this point, thus making the contribution of CityEXP less significant. Furthermore, I do not see how the rich scene information assists the learning, as only building geometry appears to be used in training.

  2. It seems that CityEXP contains only four city-scale scenes. Detailed statistics are scattered throughout the paper, without a clear comparison and discussion with existing datasets. 9,269 building instances do not seem to constitute a particularly large-scale dataset, though the paper indicates CityEXP is scalable. It is also unclear how many instances are actually used for training/validation/testing.

  3. There are no other methods (excluding wireframe experiments) evaluated on the proposed benchmark or trained with the proposed dataset. Therefore, the role of CityEXP as a benchmark is unclear.

  4. As this work focuses on urban buildings, polygonal meshes with large planar faces would be a more appropriate output representation. However, the current mesh output from UrBlend consists of dense triangular meshes (from marchingcube extraction), without good planarity over large regions or preservation of sharp features. This limitation appears to be related to the use of Hunyuan3D's 3D representation, which does not seem well suited to building geometry.

  5. It is unclear how LoD information is used for training. If all different levels of meshes are treated as GT for training, how is the detail level of the generated output controlled?

  6. UrBlend is a relatively simple extension of Hunyuan3D-Omni, and the novelty of the network design is limited. The reconstructed results also do not contain much detail, for instance, most windows are missing. A better design should utilize more structural information for encoding and decoding, not SDF alone.

  7. The protocol paragraph in Sec. 6.1 states, "We adapt Ours from the BuildingWorld3D checkpoint...". This design somewhat weakens the motivation for creating CityEXP for the overall work, as the model relies on substantial prior training on another building dataset. Furthermore, the performance using the original checkpoint should be reported as well.

  8. Lines 377/409-411: On input-supported regions, the distance increases significantly from 5.31 to 23.93 after applying the proposed method. I think this result is quite unfavorable, indicating that the method requires substantial further improvement. Note that many point completion works preserve existing geometry well, check: PoinTr, SnowflakeNet (ICCV 2021), DANCE(AAAI 2026). I would suggest enforcing the reconstructed geometry to align with the observed points, for instance, by setting SDF(p) = 0 during the sampling procedure.

  9. The experiments in Section 6.2, "Conditioning Ablations on BuildingWorld3D," do not show a significant advantage of the proposed design over the other choices.

Writing Issues

The paper contains many unclear terms and broken logical flows, making it unnecessarily difficult to read. Many passages also lose logical connections between sentences.

Here I list a few issues for reference.

  • Line 021: "compatible": its meaning is unclear. A formal definition should be provided.
  • Line 153: "conversion unverified": Who performed the conversion, and who is responsible for verifying it?
  • Line 158: Why does the table not contain a "compact" column, which would seem more relevant to building reconstruction?
  • Line 159: "Autoregressive targets must meet topology and size limits": Why does this sentence appear here? Its logical connection with the surrounding sentences is unclear.
  • Line 160: The terms "recoverable" and "validity" are vague. Appendix D.1 lists a set of criteria, but further explanations are still needed, for instance, for "coordinates".
  • Table 1 is not referenced in the paper. BuildingWorld3D and 3DBAG are not listed in Table 1, and the conclusion to be drawn from this comparison is unclear to me. The data scales of the different datasets are also not reported.
  • Figure 2 caption: Why is the term "evaluated baseline" used here? DA3 is the method used to create the point clouds, so calling it a "baseline" is confusing.
  • There is no description of the evaluation metrics. F.01/F.02 should be explained, and the choice of using only Chamfer distance should be discussed. Corner/edge accuracy should also be considered. As a benchmark protocol, evaluation metrics should be well designed, and a review of existing metrics would help this design.
  • Table 2, why "valid" is reported? As the meshes are extracted from marching cubes, they have no valid issue in general.
  • As the point clouds are described in real-world scale, please report Chamfer distance in physical units, e.g., cm or mm.
Questions for authors

It is necessary to clearly explain the role of CityEXP, provide its data statistics, and discuss the main conclusions that can be drawn from the benchmark, rather than presenting only the experimental results of UrBlend.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.