acceptodds
Under review as a conference paper at ICLR 2027

T3lescope: Arbitrary-Resolution High-Fidelity Generative Surface Reconstruction from Images

Abstract

We reconstruct high-fidelity 3D scene meshes from posed multi-view images without per-scene optimization, across scales ranging from single objects to large outdoor scenes. Per-scene optimization methods lack the learned 3D prior needed when observations are sparse or surfaces are glossy or transparent. Existing generative methods leverage such priors to complete geometry in sparsely observed regions, but typically operate at a fixed resolution over a limited spatial extent, trading spatial coverage against detail. Reconstructing a large scene therefore often requires partitioning it into independently processed overlapping local regions, making it difficult to maintain global geometric consistency. To address these issues, we propose T3lescope, which applies a single fixed-resolution generator across scene scales in an inference-time coarse-to-fine cascade. A coarse level establishes the scene layout, and finer levels perturb and denoise geometry inherited from the coarser level within progressively finer spatial cells to recover surface detail. The model is trained on individual cells at multiple scales and shares its weights across all levels, so no hierarchy is fixed during training, and the number of levels, cell scales, and cell locations are determined at inference time. On indoor, outdoor, and city-scale scenes, T3lescope outperforms feed-forward and generative baselines, matches or surpasses per-scene optimization, and recovers fine structures as well as glossy and transparent surfaces. These results show that our method generalizes across diverse scenes, view counts, and image resolutions.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D reconstruction, mesh, generative model

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_ZUSVOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper proposes T3lescope, a method that reconstructs scene meshes from posed multi-view images without per-scene optimization. It builds on TRELLIS.2 and reuses a single fixed-resolution generator across scales through an inference-time coarse-to-fine cascade. A coarse level generates the global layout. Each finer level halves the cell size and refines the geometry inherited from its parent via SDEdit. The model is trained only on individual cells at multiple scales, so the depth and scale of the hierarchy are chosen at inference time. Image conditioning combines voxel-queried attention over DINOv3 tokens, using relative ray, offset, and scale encodings, with a ConvNeXt detail branch and a cell-independent multi-resolution tile bank. The method is evaluated on ScanNet++, Tanks and Temples, and an in-house city-block dataset. It outperforms the generative baseline GenRecon, produces substantially cleaner and more complete surfaces than feed-forward and optimization baselines, and handles glossy and transparent surfaces.

Critical strengths
  1. A scalable design for generative reconstruction that holds up empirically. Reusing one generator across scales through an inference-time cascade lets a fixed-resolution 3D prior cover everything from room scale to city blocks while keeping the global structure consistent. The parent-initialization ablation supports this clearly: without the parent, CD rises from 0.0212 to 0.1051 on large ScanNet++ scenes, and roofs fragment on city blocks. On city blocks, the method reaches a CD of 0.65 m, while all baselines are at 4.9 m or worse.

  2. Clearly superior surface quality. The meshes are cleaner, sharper, and more complete than all baselines, including on glossy and transparent surfaces. The normal error confirms this, e.g., 16.3° vs. 28.0° for DA3 on large ScanNet++ scenes. Notably, the general model is trained only on synthetic data and transfers zero-shot to real captures.

Critical weaknesses
  1. Key configurations are hand-set, with no selection rule and no sensitivity analysis. The paper says that cell scales, layout, and cascade depth are "determined at inference time." In practice, they are set by hand per dataset or per scene (Appendix C.5):

    • initial cell size and number of levels: 4/2/1 m on ScanNet++, 80/40/20/10 m on city blocks, 2.8/1.4/0.7 for the human capture;
    • the normalization target for non-metric scenes: "6 or 8" on Tanks and Temples;
    • the source of the scene extent: DA3 points, the camera hull, or the GT footprint;
    • height caps.

    These hand-crafted hyperparameters will constrain the generalization ability of the method. Also, no rule is given for a new scene, no sensitivity analysis is reported, and it is unclear on which data these choices were made. In addition, t_s is fixed at 0.8 even though Appendix G.1 shows a view-density-dependent trade-off between global consistency and local accuracy. All of this weakens the claim of practical generality.

  2. The evaluation does not fully capture the method's strengths. CD and F@2 measure only positional accuracy, and on ScanNet++ they favor DA3, while the qualitative results and the normal error show much better surface quality for the proposed method. The current metrics neither quantify surface quality directly nor separate a global offset from shape error. Yet since the GT itself has imperfection which leads to more difficulties in fair metric comparisons. I am wondering if the PSNR of rendered depth / normal / pointmap would be better to evaluate the geometry as a complement.

  3. Missing discussion of and comparison with related work, and no guarantee of fidelity to the inputs.

    • Related work: Two closely related lines of work are not discussed. The first is reconstruction-guided generative reconstruction: ReconViaGen (ICLR 2026) and UniRecGen/RecGen3D (SIGGRAPH Asia 2026). The second is feed-forward volumetric scene reconstruction: Atlas, NeuralRecon (itself coarse-to-fine), and FineRecon.
    • Comparisons: GenRecon appears only on ScanNet++. Prior-regularized optimization methods (e.g., MonoSDF, DN-Splatter) are missing, although they are needed to test the paper's claim that optimization lacks learned priors. Photogrammetry is also missing, e.g., RealityScan on the city blocks, which the authors already use for registration.
    • Fidelity: The authors acknowledge that their surfaces are offset from the observations. Yet nothing in the method enforces consistency with the inputs: DA3 geometry is available but used only for cell placement and view selection. Both ReconViaGen and UniRecGen address this issue explicitly.

Minor:

  • The output is geometry only; GenRecon outputs PBR meshes.
  • The method requires posed images and fails on unbounded scenes.
  • Some training data details are missing, e.g., the size of the TexVerse subset used, and there is no commitment to release the synthetic data.
Questions for authors

Clarifications:

  1. How should s₀, the number of levels, the normalization target, and the scene extent be chosen for a new scene? How sensitive are the results to these choices, e.g., s₀ ∈ {2, 4, 8} m, or a normalization target of 6 vs. 8 on Tanks and Temples? On which data were they selected?

Requests that could change my assessment:

  1. Please consider reporting metrics that capture surface quality and separate offset from shape error: SSIM/PSNR on rendered normal maps, or F-score after local ICP alignment and at multiple thresholds, or any other applicable metrics.

  2. Can t_s be chosen adaptively, e.g., based on view density or scene extent?

Suggestions:

  1. Discuss reconstruction-guided generative reconstruction and feed-forward volumetric methods in the related work, and if feasible, compare against a prior-regularized optimization baseline and against photogrammetry on the city blocks.

  2. If possible, could the authors provide any qualitative results on any in-the-wild / real-world captured image/video sets? It would be promising to show the generalization ability of a scalable method.

5 comments
AuthorsOpenReview ↗

Response to Reviewer ZUSV (1/3)

We thank the reviewer for the detailed review and the concrete suggestions, which helped us improve the evaluation and the discussion of related work.

W1/Q1. Choosing the configuration. Only the ratio between the scene scale and the cell sizes matters, because the generator sees only quantities relative to the cell. For example, doubling the scene scale is the same as halving s0s_0 and all finer cells. We follow the rules in Appendix I.4.

  • Scene scale. We set the scene scale relative to the finest cell size. For SfM captures, we rescale the COLMAP coordinates so that the median distance from the cameras to their centroid equals a target value, and the cell sizes are given in these units (Appendix C.5). The inference cost grows with a larger target or a smaller finest cell, and a target of about 6 to 8 times the finest cell is the practical limit. All our scenes stay within this limit.
  • Coarsest cell s0s_0. The generator sees one cell at a time, with only coarse context from its neighbors through MultiDiffusion. We therefore choose s0s_0 large enough that a cell and its neighbors determine the shape inside it. For rooms, we use 4 m, about the ceiling height, so that a cell spans from the floor to the ceiling. For city blocks, we use 80 m, so that two cells cover a building.
  • Number of levels. Levels can be added as long as the computational budget and the image resolution allow. Each finer level halves the cell size and increases the runtime by about 5 to 8 times. We use three levels for rooms and four for city blocks as a practical trade-off between detail and runtime.
  • Sensitivity. On Truck and Caterpillar, we lower the target from 6, the value used for these two scenes, to 4 and keep all other settings, which is the same as making the cells at every level, including s0s_0, 1.5 times larger. The lower target reduces the mean F@τ in all six input settings, while the Chamfer distance changes in both directions (Table 19). Lowering the target thus reduces the accuracy of fine details consistently but only slightly.
  • A new scene. These settings can be set without the ground truth. We choose the number of levels from the desired surface resolution and the computational budget, choose s0s_0 from the size of the structures to reconstruct, such as the room height or the size of a building, and normalize SfM poses so that the median camera distance is at most about 6 to 8 times the finest cell. Both the height range and the horizontal extent of the coarsest cells can be set by hand or computed from the DA3 points, and the horizontal extent can also be computed from the convex hull of the camera trajectory (Appendix C.5).
  • How the settings were selected. We use the following settings (Appendix C.5).
    • ScanNet++: cells of 4, 2, and 1 m, so that the finest voxels are about 2 mm. The horizontal extent and the height range are computed from the DA3 points.
    • City blocks: cells of 80, 40, 20, and 10 m, so that the finest voxels are about 2 cm. The horizontal extent is the footprint of the ground-truth mesh, because the evaluation crops all reconstructions to this footprint (Appendix B) and the DA3 points of these unbounded scenes extend far beyond the block. The height range is computed from the DA3 points.
    • Tanks and Temples: cells of 4, 2, and 1 in the normalized units. The target is 6 on Ignatius, Truck, and Caterpillar, which are captured by orbiting an object, and 8 on Barn, Courthouse, and Meetingroom, which have a larger space of interest. The horizontal extent is computed from the DA3 points on Meetingroom and from the convex hull of the camera trajectory on the other scenes.
    • Human capture: cells of 2.8, 1.4, and 0.7 with a target of 4, and the height range is capped at 4.

W1/R2. Adaptive tst_s. We analyzed the relationship between the number of images observing a cell and the tst_s that gives the higher accuracy in that cell (Appendix I.3). At the finest level, ts=1t_s = 1 gives a higher per-cell F@2 for cells observed by 15 or more images, whereas ts=0.8t_s = 0.8 is better for cells observed by fewer than 10 images (Figure 30). This per-cell F@2, however, measures only the local accuracy. On whole scenes, ts=0.8t_s = 0.8 gives the lowest Chamfer distance and the highest NC in all settings (Table 10), because the parent prevents hallucinated surfaces, especially with sparse views (Table 18). Choosing tst_s per cell while keeping this consistency, that is, balancing reconstruction and generation according to the observations, is an important direction for future work.

AuthorsOpenReview ↗

Response to Reviewer ZUSV (2/3)

W2/R1. Metrics for surface quality. We thank the reviewer for suggesting these metrics. We agree that CD and F@2 measure only the positional accuracy and do not separate local offsets from shape errors, and we added the suggested metrics.

  • Table 2 reports the angular error, the fractions of pixels within 11.25° and 30°, and the PSNR, SSIM, and LPIPS of rendered normal maps, and Figure 4 shows the normal maps. With both 32 and 256 views, our method is the best on all of these metrics except SSIM.
  • Appendix A.1 reports the PSNR, SSIM, and LPIPS of rendered depth maps (Table 6). DN-Splatter + DA3, which fits each scene to the DA3 depth, has the highest PSNR and SSIM, followed by our L1 and L2, while our method has the lowest LPIPS in both settings.
  • Appendix I.1 reports F-scores at multiple thresholds and after local ICP in each 25 cm cube, which separates local offsets from the shape (Table 17). After this alignment, the gap in F@2 to DN-Splatter + DA3 shrinks from 0.06 to 0.004 with 32 views and from 0.08 to 0.02 with 256 views, and our Chamfer distance is the lowest among the methods without per-scene optimization. The offsets of our cubes are comparable to those of DA3 and DN-Splatter + DA3 and about half of those of GenRecon.

The rendered normal and depth maps of all methods, including the added baselines (Figure 4 and Figures 13 to 16), show that our surfaces have flat faces and sharp edges. They also show that the ScanNet++ ground truth itself is noisy, so the small remaining gaps do not necessarily indicate a less accurate surface shape.

W3. Related work. Section 2 now discusses both lines of work. Reconstruction-guided methods such as ReconViaGen and RecGen3D condition an object-level generator on a feed-forward reconstruction of the input views, and ReconViaGen further guides the late denoising steps with a rendering loss against the inputs. They reconstruct a single object in one normalized volume at the generator's fixed resolution, whereas our method extends a fixed-resolution generator to whole scenes through a cascade of cells. Feed-forward volumetric methods such as Atlas, NeuralRecon, and FineRecon predict the geometry directly in 3D, as ours does, and NeuralRecon also proceeds from coarse to fine. However, they regress the geometry from image features, which tends to smooth out fine details. Our method instead samples the geometry from a generative 3D prior and recovers finer details than FineRecon + DA3 in our comparison (Figure 4).

W3. Comparisons. On ScanNet++, we added MonoSDF, DN-Splatter, DN-Splatter + DA3, and FineRecon + DA3 (Tables 1 and 2), and on the city blocks, DN-Splatter + DA3 (Table 3(b)). The methods that use the DA3 depth have a higher F@2 than ours, and DN-Splatter + DA3 also has a lower Chamfer distance. The gap comes mostly from small local offsets and largely closes after local ICP (see W2/R1), and our method gives the most accurate normals. On the city blocks, where the DA3 depth is unreliable, our method obtains a Chamfer distance of 0.61 m, compared with over 3 m for all baselines including DN-Splatter + DA3.

W3. Discussion on enforcing consistency with the inputs. Our method does not explicitly enforce consistency with the input images, and uses DA3 only to place the cells and to select the views, for two reasons (Appendix I.2). First, enforcing consistency in the output space is too costly for scenes. ReconViaGen renders the decoded latent and corrects it with the error against the input images, which works for a single object. A cell in a scene, however, may be occluded by other cells, so rendering it requires decoding its neighbors as well. We therefore improve fidelity in the latent space by conditioning each voxel on the image features at its projection. Second, DA3 depth is not accurate enough on large outdoor scenes. Its Chamfer distance is about 2 cm on ScanNet++ but above 5 m on the city blocks, and its depths are inconsistent across views (Figure 29). Conditioning directly on the DA3 points would therefore carry these errors into our surfaces. Even without explicit consistency, our method achieves the lowest Chamfer distance after local ICP among the methods without per-scene optimization (Table 17). On scenes where the reconstructed points are accurate, such as ScanNet++, conditioning each cell on the points inside it would fit naturally into our framework, and we leave this for future work.

AuthorsOpenReview ↗

Response to Reviewer ZUSV (3/3)

Minor.

  • Geometry only. In this work, we focused on geometry because our goal was accurate surface reconstruction. The framework can also generate textures. We trained the same generator to predict the latents of the TRELLIS.2 material VAE and ran it on the garden scene from Mip-NeRF 360. It produces PBR materials for the whole scene, which also allow relighting (Appendix I.8, Figure 34).
  • Posed images and unbounded scenes. Our method currently requires posed images and does not handle distant regions of unbounded scenes (Appendix A.5). Extending it to unposed images and unbounded scenes is future work.
  • Training data. We used 108,791 TexVerse objects for pretraining and will release the code that generates the synthetic houses and grass scenes.

In-the-wild captures. In addition to the captures in Figures 1 and 3, we reconstructed three DL3DV scenes with the same model, without any training on DL3DV or per-scene optimization (Appendix I.7, Figure 33). Our method reconstructs clean and detailed surfaces on these real-world captures.

For W3. GenRecon and photogrammetry on the city blocks, we are running these comparisons and will report the results in a follow-up comment.

We hope these results address the reviewer's concerns.

Reviewer_A5irOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a framework for high-fidelity 3D mesh reconstruction from posed multi-view images without per scene optimization. The main idea is an inference-time coarse-to-fine cascade that reuses a fixed resolution 3D generator across different scene scales. The coarse level establishes the global scene structure, while finer levels refine the inherited geometry to recover surface details. The method also introduces multi-scale image conditioning to provide image features at different resolutions. Experiments on datasets ScanNet++, Tanks and Temples, and city-scale scenes show improvements over feed-forward and generative baselines. It also shows competitive results against per scene optimization methods.

Critical strengths
  1. The method reuses the same generator across multiple scene scales, it makes the method flexible to do coarse-to-fine reconstruction at inference time.

  2. The method generates high-resolution meshes without per-scene optimization. This makes it applicable to different scenes without scene specific training.

  3. The method achieves good results on indoor, outdoor, and city-scale scenes. It also recovers fine structures and challenging glossy and transparent surfaces. These results demonstrate its effectiveness across different scene types.

Critical weaknesses
  1. In the city-scale experiments, the proposed method uses images at 9504 × 6336 resolution. The baseline methods use downsampled images at 1920 × 1280 resolution. This difference may contribute to the final performance gap. Is it possible to use the same image resolution to do the comparisons? Because the current experimental settings make the comparison less convincing.

  2. For the coarse-to-fine cascade inference, each inference level uses the geometry predicted by the previous level for initialization and cell placement. Therefore, errors or missing structures at the coarse level may affect the reconstruction at finer levels. Is there any method to detect or analyze how such errors propagate across different cascade depths?

  3. In the method, the generator does not receive the physical cell size or voxel spacing as input, it only relies on multi-scale image features and projected voxel size. So it is unclear whether this design can reliably generalize to scene scales beyond those seen during training. Could the authors give some discussions about it?

Questions for authors
  1. Is it possible to use the same image resolution to do the comparison in the city-scale experiments?

  2. For the coarse-to-fine cascade inference, is there any method to detect or analyze how such errors propagate across different cascade depths?

  3. In the method, if the generator does not receive the physical cell size or voxel spacing as input, how can the method be reliably generalized to scene scales beyond those seen during training? Could the authors give some discussions about it?

3 comments
AuthorsOpenReview ↗

Response to Reviewer A5ir

We thank the reviewer for the constructive feedback and for recognizing the flexibility of reusing one generator across scales and the results across indoor, outdoor, and city-scale scenes.

W1/Q1. Same image resolution on the city blocks. We now evaluate all methods with the same 1920×1280 images (Table 3(b) and Table 9). At this resolution, our method still outperforms all baselines by a large margin in all metrics, reaching a Chamfer distance of 0.61 m compared with over 3 m for the baselines.

W2/Q2. Error propagation across cascade levels. Appendix I.3 analyzes how errors propagate across the cascade levels. On five large ScanNet++ scenes, we follow every ground-truth point through L0, L1, and L2 and classify it at each level as accurate (within 2 cm, as in F@2) or an error. Missing structures are included, since a ground-truth point without a prediction nearby counts as an error. Between two levels, we then count the inherited errors (errors that remain errors), the corrected errors, and the new errors (accurate points that become errors). We also measure the false surface at L2, the fraction of predicted points at least 10 cm from the ground truth (Table 18).

With the parent initialization (ts=0.8t_s = 0.8), most errors at a coarser level are inherited (65% with 32 views and 72% with 256 views from L0 to L2), but only about 5% of the accurate points become new errors. Without the parent (ts=1t_s = 1), where the finer levels start from noise and inherit only the cell placement, they correct more errors but also introduce more new ones. With 32 views, they hallucinate many surfaces, and 33.88% of the predicted points at L2 are false surface, compared with 1.78% with the parent. With 256 views, the dense observations constrain the generation enough that the finer levels correct most errors of the coarser level without the parent, which likely explains why ts=1t_s = 1 reaches a higher F@2 in several settings of Table 10. The parent still gives the lowest Chamfer distance and the highest NC in all settings, so it mainly prevents hallucinated surfaces, especially when the views are sparse. Balancing reconstruction and generation according to the available observations, for example by adapting the parent initialization per cell, is an important direction for future work. Please see Appendix I.3 of the revised paper for the full analysis.

W3/Q3. Generalization to unseen scales. The generator never receives the physical scale, so the scale of a cell reaches it only through how the cell appears in the images. All of its inputs are relative to the cell. The image features are sampled at the resolution that matches the voxel spacing, and the projected voxel size is encoded relative to the image. A small cell and a large cell therefore look the same to the generator when their images show the same structure at the same relative resolution (Appendix I.4). Accordingly, our method reconstructs clean and detailed surfaces on captures whose scales differ from those in training, including Tanks and Temples (Figure 3(b)), garden from Mip-NeRF 360 (Figure 1), our capture of five people (Figure 3(c)), and three DL3DV scenes added in Appendix I.7. The model fails on distant regions of unbounded scenes (Appendix A.5), which we attribute to biased viewpoints rather than scale (Appendix I.4).

We hope these results address the reviewer's concerns.

Thanks for the response. My concerns have been addressed, and I will raise my rating to weak accept.

AuthorsOpenReview ↗

Reply to Reviewer A5ir

We sincerely thank you for increasing the rating, and we are glad that your concerns have been addressed.

Reviewer_Zeb3OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents T3lescope, a framework that extends a fixed-resolution TRELLIS.2 generator to high-resolution scene surface reconstruction through a coarse-to-fine cascade. The method combines joint denoising over overlapping cells, parent-guided sparse structure initialization, and multi-scale image conditioning using a shared tile feature bank and voxel-queried attention. Experiments on ScanNet++, Tanks and Temples, and city blocks demonstrate improvements over existing generative reconstruction methods and competitive results against feed-forward and per-scene optimization approaches.

Critical strengths
  1. The reconstruction results are strong overall, with substantial improvements over the existing generative baseline. The method also demonstrates encouraging recovery of scene structure and fine details under sparse observations and on challenging surfaces.
  2. The experimental evaluation is extensive, covering multiple datasets, scene scales, and input-view counts. Component ablations and qualitative comparisons provide useful evidence for the effectiveness of the framework.
  3. The paper is clearly written and easy to follow, with a well-organized presentation of the method and experiments.
Critical weaknesses
  1. The effect of the generative prior on geometric accuracy requires further analysis. In several settings of Table 2(a), particularly with dense observations, the method obtains lower F-scores than per-scene optimization methods, despite competitive Chamfer distances. The authors attribute this to small surface offsets, but it remains unclear whether the generative prior alters local structures constrained by the observations. The lower F-scores with parent initialization in some settings of Appendix Table 7 also suggest that inherited coarse-level errors can limit fine-level accuracy.
  2. The independent benefit of voxel-queried attention is not fully established. In Figure 6, replacing the GenRecon-style conditioner with voxel-queried attention yields only a small improvement, while the larger gains appear after adding the ConvNeXt branch. A comparison between the original conditioner and the proposed attention with the same ConvNeXt detail features would help distinguish the contributions of the attention mechanism and the additional image features.
  3. High-resolution reconstruction remains computationally expensive. Although the method avoids per-scene optimization, reconstruction through L2 takes approximately 89 minutes per scene on a single H200, with peak GPU memory of 79 GiB, in the reported large-scene ScanNet++ benchmark. This cost limits the practical accessibility of the proposed resolution scaling.
Questions for authors
  1. GenRecon also supports texture generation. Why does this work focus on geometry without continuing to generate textured results? Can the proposed multi-scale framework be extended directly to texture generation, or would consistency across cells and scales require additional design?
  2. Appendix F.1 merges boundary vertices using their integer voxel coordinates. If adjacent blocks predict different positions for the same boundary vertex, how is its final position determined? Close-up visualizations or measurements of boundary position and normal discrepancies would help verify the geometric quality of the merged mesh.
  3. Why does GaussianWrapping perform relatively poorly on ScanNet++? These results appear weaker than I would expect from its original paper. Could the authors analyze whether the input-view count, training configuration, mesh extraction, or evaluation protocol contributes to this discrepancy?
4 comments
AuthorsOpenReview ↗

Response to Reviewer Zeb3 (1/2)

We thank the reviewer for the careful review and for recognizing the strong results over the generative baseline and the extensive evaluation.

W1. Effect of the generative prior on geometric accuracy. We separated small offsets from changes in the surface shape in Appendix I.1.

  • ScanNet++. We align the prediction to the ground truth by local ICP in each 25 cm cube, which removes local offsets but keeps the shape within the cube. After this alignment, the gap in F@2 to DN-Splatter + DA3 shrinks from 0.06 to 0.004 with 32 views and from 0.08 to 0.02 with 256 views (Table 17). The mean translation and rotation of our cubes are comparable to those of DA3 and DN-Splatter + DA3 and about half of those of GenRecon. Together with our more accurate normals (Table 2), this indicates that the lower F-score mostly reflects small offsets rather than local structures altered by the prior.
  • Tanks and Temples. F@τ counts only the points within the official per-scene threshold τ (3 to 25 mm), whereas the Chamfer distance averages all distances. Figure 28 plots, for each distance on the horizontal axis (in multiples of τ), the percentage of predicted points whose distance to the ground truth is below it. The value at τ is thus the precision used in F@τ, and a curve that reaches 100% only at large distances indicates many points with large errors. On the large scenes with 256 views, more than 10% of the points of 2DGS and GaussianWrapping lie farther than 100τ from the ground truth, which inflates their Chamfer distance, while fewer than 2% of ours do. Our surfaces thus have small offsets but rarely produce large failures such as floaters, so our Chamfer distance is the lowest in most settings although our F@τ is lower. On the small scenes, per-scene optimization places more points within τ, and all methods have few points far from the ground truth.

W1. Parent initialization. To analyze how errors at a coarser level affect the finer levels, Appendix I.3 follows every ground-truth point through the cascade levels on five large ScanNet++ scenes (Table 18). With the parent, most errors at a coarser level are inherited, but fewer new errors are introduced than without it. Without the parent (ts=1t_s = 1), where the finer levels start from noise and inherit only the cell placement, they correct more errors but also introduce more new ones. With 32 views, 33.88% of the predicted points at L2 lie at least 10 cm from the ground truth, compared with 1.78% with the parent. With denser views, the observations constrain the generation enough that ts=1t_s = 1 obtains a higher F@2 in three of the four settings of Table 10 (all but the large scenes with 32 views), but the parent still gives the lowest Chamfer distance and the highest NC in all settings. The parent thus reduces hallucinated surfaces, especially when the views are sparse. A per-cell analysis (Figure 30) shows that ts=1t_s = 1 becomes better for cells observed by 15 or more images, so adapting tst_s per cell is a promising direction.

W2. Independent benefit of voxel-queried attention. We added a variant of the GenRecon-style conditioner that uses identical ConvNeXt features as our model (Appendix I.5, Table 20), on Tanks and Temples with 32 views and 4K inputs. With identical ConvNeXt features, voxel-queried attention improves the mean F@τ over the GenRecon-style conditioner by 0.024, 0.042, and 0.038 at L0, L1, and L2, with a lower CD/τ at L0 and L1 and a comparable one at L2. It also recovers fine parts such as ears, eyes, and fingers in more detail (Figure 31). The ConvNeXt features thus help voxel-queried attention much more than the GenRecon-style conditioner, so voxel-queried attention is effective for recovering detailed geometry.

W3. Computational cost. The computational cost is a current limitation of our method, and most of the runtime is spent on flow sampling (Appendix H). Speeding up the reconstruction while keeping fine details is an important direction for future work, for example by distilling the flow models into few-step samplers and by refining only the cells that need finer detail.

AuthorsOpenReview ↗

Response to Reviewer Zeb3 (2/2)

Q1. Texture generation. In this work, we focused on geometry because our goal was accurate surface reconstruction. The framework can also generate textures. We trained the same generator to predict the latents of the TRELLIS.2 material VAE and ran it on the garden scene from Mip-NeRF 360. It produces PBR materials for the whole scene, which also allow relighting (Appendix I.8, Figure 34).

Q2. Boundary vertices. We added close-ups and measurements of the seams between adjacent blocks on a ScanNet++ scene (Appendix I.6, Figure 32). Each block is decoded with a 4-voxel latent margin that overlaps its neighbor, so both blocks predict every seam vertex. The two predictions are almost identical, differing by less than one voxel at every seam vertex (median 5.5×10⁻⁴ voxels), and the normals of the same face differ by 0.017° at the median. When the two predictions differ, we keep the position predicted by the block that contains the vertex and connect the neighbor's faces to it, so the two sides share a single vertex and no gap opens.

Q3. GaussianWrapping on ScanNet++. GaussianWrapping performed worse with the full 1752 px images used in the submission than with half-resolution images (876 px), so we replaced its results with the half-resolution ones (Appendix A.1, Table 7). Our method still outperforms it on all metrics of Table 1.

We hope these results address the reviewer's concerns.

Thank you for the thoughtful response. My questions have been largely addressed, and therefore I will maintain my positive score.

AuthorsOpenReview ↗

Reply to Reviewer Zeb3

We sincerely thank you for your constructive feedback and positive evaluation, and we are glad that your questions have been addressed.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response

General response

We thank all reviewers for their thoughtful and constructive feedback. We are encouraged that the reviewers found reusing one generator across scales a scalable and flexible design (ZUSV, A5ir), the results strong with substantial gains over the generative baseline (Zeb3) and effective across indoor, outdoor, and city-scale scenes (A5ir, ZUSV), the surface quality clearly superior (ZUSV), and the paper clear and easy to follow (Zeb3, A5ir, ZUSV). We have uploaded a revised paper, in which added or changed text is shown in blue and most of the new analyses are collected in Appendix I. Below we summarize the revisions and the comparison with the added baselines. The other questions are answered in the response to each reviewer.

Summary of revisions

  1. Additional baselines (ZUSV, Zeb3). On ScanNet++, we add prior-regularized per-scene optimization methods (MonoSDF and DN-Splatter) and a feed-forward volumetric method (FineRecon) to Tables 1 and 2. DN-Splatter is evaluated both with its default depth prior and with the DA3 depth that our method uses for cell placement (DN-Splatter + DA3), and FineRecon takes the same DA3 depth (FineRecon + DA3). GaussianWrapping performs better with half-resolution images (876 px) than with full-resolution ones, so we now report it at half resolution (Appendix A.1, Table 7). Appendix A.1 also compares the added baselines qualitatively, with the meshes in Figures 8 to 11 and the rendered normal and depth maps with their error maps in Figures 13 to 16.
  2. Surface-quality metrics (ZUSV). Table 2 reports the angular error, the fractions of pixels within 11.25° and 30°, and the PSNR, SSIM, and LPIPS of rendered normal maps, and Figure 4 shows the normal maps. Appendix A.1 adds the same image metrics for depth maps (Table 6), and Appendix I.1 reports F-scores at multiple thresholds and after local ICP alignment (Table 17).
  3. City blocks at the same input resolution (A5ir). All methods, including ours, now take the same 1920×1280 images (Table 3(b) and Table 9), and DN-Splatter + DA3 is added.
  4. Related work (ZUSV). Section 2 now discusses reconstruction-guided generative reconstruction (ReconViaGen, RecGen3D) and feed-forward volumetric reconstruction (Atlas, NeuralRecon, FineRecon).
  5. Additional analyses (Appendix I).
    • I.1 Local offsets and error distribution (Zeb3, ZUSV)
    • I.2 Discussion on enforcing consistency with the inputs (ZUSV)
    • I.3 Parent initialization across cascade levels, covering error propagation and adaptive tst_s (Zeb3, A5ir, ZUSV)
    • I.4 Configuration choices and unseen scales (ZUSV, A5ir)
    • I.5 Conditioner ablation with the same ConvNeXt features (Zeb3)
    • I.6 Boundary vertices between adjacent blocks (Zeb3)
    • I.7 In-the-wild captures (ZUSV)
    • I.8 Material generation (Zeb3, ZUSV)

Comparison with the added baselines

Our method produces the most accurate surface normals. With both 32 and 256 views, it has the highest NC in Table 1, is the best on all normal-map metrics in Table 2 except SSIM, and gives visibly cleaner surfaces (Figure 4). In Table 1, the methods that use the DA3 depth have a higher F@2 than ours, and DN-Splatter + DA3 also has a lower Chamfer distance in both view settings. This gap comes mostly from small local offsets. After local ICP alignment, the gap in F@2 to DN-Splatter + DA3 shrinks from 0.06 to 0.004 with 32 views and from 0.08 to 0.02 with 256 views, and our Chamfer distance is the lowest among the methods without per-scene optimization (Table 17). Since the ScanNet++ ground truth is itself noisy (Figure 4), the small remaining gap does not necessarily indicate a less accurate surface shape. On the city blocks, where the DA3 depth is unreliable, our method obtains a Chamfer distance of 0.61 m with the same 1920×1280 images as the baselines, whereas all baselines exceed 3 m (Table 3(b)). Overall, our method achieves the highest surface quality among the methods without per-scene optimization and is competitive with prior-regularized per-scene optimization.

Note on numbering

Adding tables and figures has shifted some numbers. The tables and figures referred to in the reviews correspond to the revision as follows. The section numbers of the appendix are unchanged.

In the reviewsIn the revision
Table 2(a)Table 3(a)
Table 2(b)Table 3(b)
Angular error in Table 1Table 2
Figure 5Figure 6
Figure 6Figure 7
Appendix Table 7Table 10

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.