- 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
-
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.
-
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
-
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.
-
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.
-
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:
- 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:
-
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.
-
Can t_s be chosen adaptively, e.g., based on view density or scene extent?
Suggestions:
-
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.
-
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
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.
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.
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 s0 and all finer cells. We follow the rules in Appendix I.4.
W1/R2. Adaptive ts. We analyzed the relationship between the number of images observing a cell and the ts that gives the higher accuracy in that cell (Appendix I.3). At the finest level, ts=1 gives a higher per-cell F@2 for cells observed by 15 or more images, whereas ts=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=0.8 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 ts per cell while keeping this consistency, that is, balancing reconstruction and generation according to the observations, is an important direction for future work.