acceptodds
Under review as a conference paper at ICLR 2027

GenAnchor: Observation-Anchored Progressive Reconstruction with Generative Priors for Sparse-View Gaussian Splatting

16 views

Abstract

Recent sparse-view 3D reconstruction methods exploit generative priors to com- plete under-observed scene content. However, generated views are not necessarily grounded in the available input evidence and may therefore conflict with real ob- servations, degrading the appearance and geometry of reliably reconstructed re- gions. To address this issue, we present GenAnchor, an observation-anchored progressive reconstruction framework that preserves reliably observed content while selectively applying generative supervision to uncertain and unobserved re- gions. Specifically, we first employ coverage-aware trajectory planning to explore inadequately reconstructed regions. Each planned view is then partitioned into preservation, refinement, and generation regions using a tri-state masking strat- egy, thereby adapting generative supervision to the reconstruction status of each region. Subsequently, we recover depth from the repaired RGB sequence using the planned camera poses and depth anchors from real observations, preserving observed geometry while guiding depth inference in generated regions. The re- sulting RGB-D views are integrated into the global scene representation, whose updated state guides the next round of trajectory planning and view repair, thereby forming a progressive reconstruction loop. Extensive experiments on multiple widely used public datasets demonstrate that our method consistently improves scene completeness while better preserving reliably observed regions, achieving state-of-the-art performance in both appearance fidelity and geometric accuracy.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Scene Reconstruction, Sparse-View Reconstruction, Gaussian Splatting, Generative Priors

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_MFtSOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

Generating views in sparse-view scene reconstruction is not reliable because the generated content may conflict with the original observation. This paper GenAnchor, solves this issue by preserving reliably observed content while selectively applying the generative info to uncertain and occluded regions. GenAnchor surpasses other baseline methods on various public datasets.

Critical strengths
  1. This paper identifies a real failure case of generative reconstruction and solves it directly. Though reconstruction via generation solution is not new, this paper proposes a clear tri-state idea, which is a neutral response.

  2. Re-estimating geometry after generative repair is conceptually strong. Once diffusion changes object boundaries or generates new content, the old gs depth is actually no longer valid. This paper recognises this and anchors the depth to the real observation.

  3. They do more exps rather than simply report global psnr at novel views. The observed/unobserved region analysis is valuable, and it meets the core claims in the paper: improve the missing areas w/o destroying observed areas.

Critical weaknesses
  1. It is hard to fully prove that the attribute gains to GenAnchor because the representations and backbones varies across baselines.

  2. The choice of 2dgs is insufficiently justified and limits evidence of generality. This paper proposes a framework-level contribution, demonstrating the same idea on vanilla 3dgs and its variations would make the contributions substantially stronger.

  3. The exact integration with WorldStereo is not fully transparent from the main content and the appendix. This paper gives the formulations for masking and anchoring operations, but it is not fully clear where exactly does this modified conditioning enter the pretrained pipeline.

  4. The camera planner considers whether content can be rendered, rather than the content is verified by real observations. This may allow trajectories with substantial provisional content but little grounded content. This seems somewhat disconnected from the paper's observation-anchored motivation.

Questions for authors
  1. Why the full pipeline is based on 2dgs rather than vanilla 3dgs?

  2. How much of the performance gain is from the pipeline itself vs the choice of the pretrained models?

  3. How much computational overhead does this pipeline introduce?

  4. Will the pipeline crash if the generative models introduce more noise rather than correct and precise content?

5 comments

[W1&W2/Q1&Q2] Generality of the Framework

We appreciate the reviewer’s comment. We use 2DGS because its oriented planar primitives can explicitly represent local surfaces and provide depth and normals for geometric supervision and surface reconstruction. These properties suit our joint appearance and geometry objectives, and the comparison results below show that 2DGS achieves the strongest overall performance among the tested representations. Nevertheless, the framework's observation verification, view repair, and supervision rules operate on rendered RGB-D and real-input correspondences, allowing them to transfer across Gaussian representations.

We have completed full progressive reconstruction using two alternative Gaussian-based reconstruction backbones: vanilla 3DGS and FSGS. Both substitutions cover all eight Replica scenes. The results are summarized below, together with the baselines reported in the paper.

Table. Representation substitutions and baseline comparisons on Replica.

MethodPSNR ↑SSIM ↑LPIPS ↓CD ↓F-Score ↑NC ↑
2DGS14.970.6520.4170.03960.2400.587
FSGS19.320.7660.3350.01930.4340.579
WorldStereo14.340.6380.4630.03890.2270.584
GuidedVD22.300.8000.2750.01990.3620.710
G4Splat23.120.8260.2180.01040.7340.816
GenAnchor (3DGS)24.950.8460.2030.00580.8620.868
GenAnchor (FSGS)24.830.8480.2310.00660.8250.862
Ours25.440.8560.1850.00560.8650.876

Both substitutions outperform the compared baselines in PSNR, SSIM, CD, F-Score, and NC, demonstrating that the benefits of our framework extend beyond the original 2DGS implementation. These results support the applicability of our framework across different Gaussian-based reconstruction methods. The corresponding experiments have been added to Appendix A.11 of the revised version.

To evaluate generality across generative priors and assess the contribution of the pipeline itself, we test our method and G4Splat with ViewCrafter and WorldStereo on all eight Replica scenes, allowing the two methods to be compared under the same pretrained generative prior. Each variant retains its own reconstruction pipeline; Ours uses WorldStereo by default.

Table. Methods integrated with different diffusion models.

MethodCD ↓F-Score ↑NC ↑PSNR ↑SSIM ↑LPIPS ↓
WorldStereo0.03890.2270.58414.340.6380.463
GuidedVD0.01990.3620.71022.300.8000.275
G4Splat0.01040.7340.81623.120.8260.218
G4Splat (w. ViewCrafter)0.00950.7740.85220.110.8000.299
Ours (w. ViewCrafter)0.00710.8230.85123.270.8380.201
G4Splat (w. WorldStereo)0.00610.8560.88324.500.8560.214
Ours0.00560.8650.87625.440.8560.185

As shown in the table, our method achieves higher PSNR and F-Score and lower LPIPS and CD than G4Splat with both ViewCrafter and WorldStereo. The PSNR gains are 3.16 dB and 0.94 dB, respectively, with corresponding CD reductions of 25.1% and 7.6%.

These results are consistent with how our framework uses real observations to guide generation and reconstruction. Our tri-state masking preserves verified input appearance in Grounded regions, refines uncertain renderings in Provisional regions, and generates missing content in Uncovered regions. Our observation-anchored geometric completion preserves verified input depths and uses them to calibrate recovered geometry before it supervises Gaussian reconstruction. Our coverage-aware camera exploration targets uncovered regions adjacent to rendered context, enabling progressive scene expansion. Together, these designs direct generation toward content requiring completion or refinement and convert repaired views into reconstruction supervision constrained by verified input geometry. These consistent improvements demonstrate that the performance advantages stem from our framework design rather than from the choice of pretrained diffusion model. These results have been added to Appendix A.12 of the revised version.

[W3] Exact Integration with Pretrained WorldStereo

We thank the reviewer for pointing out the need for further implementation details. We use the pretrained four-step WorldStereo Memory-DMD pipeline based on Wan2.1-I2V-14B, without introducing additional learnable layers or fine-tuning the pretrained model. We have revised Appendix C.2 to clarify the integration as follows.

Condition construction and integration.

The tri-state partition is incorporated through multiple conditions rather than a single three-class mask. The conditional RGB sequence IcondI^{\mathrm{cond}} (render_video) contains projected real-input colors in Grounded regions, rendered colors in Provisional regions, and black pixels in Uncovered regions. The binary editing mask E=MP+MUE=M^P+M^U (render_mask) specifies editable regions, while MUM^U (render_rgb_hole_mask) controls an additional feature-gating operation.

The conditional RGB, editing mask, and camera information are incorporated through WorldStereo's pretrained conditioning pathways. We additionally introduce feature gating to suppress unreliable rendering features in Uncovered regions, while retaining the original camera and reference-image conditioning mechanisms. The resulting conditions guide the pretrained denoising Transformer without modifying its learnable parameters.

Observation-based latent replacement.

After each scheduler update, we apply Eq. (A5):

zj+1=Mz⊙Qsj+1(zref,ϵ)+(1−Mz)⊙zˉj+1.z_{j+1} = M_z \odot \mathcal{Q}_{s_{j+1}}(z^{\mathrm{ref}}, \epsilon) + (1-M_z) \odot \bar{z}_{j+1}.

Here, MzM_z selects latent cells whose pooling regions lie entirely within Grounded, and zrefz^{\mathrm{ref}} is encoded from a color-extended Grounded reference image, distinct from the original real-image references. The replacement uses reference latents at the corresponding noise level, with clean reference latents in the final step. The updated latent is passed to the next denoising step, and exact projected real-input colors are restored within the original Grounded mask after decoding.

These modifications operate during inference, while all pretrained model weights remain unchanged.


[W4] Camera Planning and Observation Anchoring

Thank you for raising this point. Camera planning and observation anchoring serve different purposes in the progressive reconstruction loop. Planning aims to expand scene coverage by directing the next trajectory toward regions that are uncovered in the current reconstruction. Rendering support is therefore the relevant planning signal: it identifies which parts of a candidate view are already represented and whether its uncovered regions are adjacent to sufficient rendered context for repair. Using observation anchoring as the trajectory-selection criterion would instead favor views with greater overlap with content directly verified by the input images, without directly measuring whether a trajectory exposes currently uncovered scene content. The planner therefore deliberately optimizes coverage expansion rather than the proportion of directly verified pixels.

By contrast, observation anchoring determines which content from a selected view can serve as reliable evidence during repair and scene optimization. Real-input correspondences define the Grounded regions, which receive observation-based protection; Provisional regions remain editable, while Uncovered regions are generated. Direct generated-view supervision is restricted to the latter two regions, and real inputs remain the only source of verification and depth anchoring. Thus, trajectories containing substantial Provisional content but limited Grounded content can still contribute to scene exploration without treating uncertain renderings as verified observations. In this way, coverage-aware planning and observation anchoring serve complementary purposes, enabling progressive scene expansion while preserving the distinction between real observations and uncertain content.

[Q3] Computational Cost and Baseline Comparisons

We appreciate the reviewer's comments on computational cost. We report the mean training time and peak GPU memory across the eight Replica scenes, measured on NVIDIA RTX 5090 GPUs. Training time includes input preparation, initialization, diffusion generation and repair, and optimization. Times are in minutes. The results are summarized below.

Table. Computational cost on Replica.

MethodMean training (min)Mean peak GPU memory (GB)
Difix3D+51.9710.12
ArtiFixer105.9024.01
WorldStereo28.8911.12
GuidedVD88.0232.04
G4Splat66.4732.31
Ours67.3417.04

Compared with G4Splat, our method achieves higher mean PSNR with a comparable training time (67.34 vs. 66.47 minutes) while reducing mean peak GPU memory from 32.31 GB to 17.04 GB (47.3%). Our method also requires less training time and peak GPU memory than ArtiFixer and GuidedVD. Overall, these results demonstrate that our method achieves better reconstruction quality with comparable or shorter training time and substantially lower GPU memory usage than these baselines. These efficiency results have been added to Appendix A.10 of the revised version.


[Q4] Robustness to Inaccurate Generation

We thank the reviewer for raising this concern. Our framework incorporates several mechanisms to mitigate errors introduced by inaccurate generation. During RGB repair, Grounded regions are protected by real-input colors and excluded from direct generated-view supervision. During geometric recovery, generated depth and Gaussian seeds are screened for validity and confidence to reduce the influence of unreliable geometry. During scene optimization, generated RGB receives a lower supervision weight than real observations, while a fixed real-view sampling probability maintains the contribution of real-input supervision throughout progressive reconstruction. Together, these mechanisms help limit the propagation of errors from generated content.

To evaluate robustness under persistent generation errors, we apply smooth random spatial deformations to generated RGB at every reconstruction round. The perturbations affect only Provisional and Uncovered regions, while Grounded regions and real inputs remain protected. The subsequent depth recovery, scene optimization, and camera planning are performed using the perturbed results. We evaluate the complete pipeline on Replica scan3 under identical reconstruction settings, with Clean denoting no additional perturbation.

Table. Final reconstruction results on Replica scan3.

ConditionPSNR ↑SSIM ↑LPIPS ↓AbsRel ↓RMSE ↓CD ↓F-Score ↑NC ↑
Strong26.130.8690.1740.0230.1250.00750.7960.856
Moderate26.580.8720.1690.0210.1120.00710.8240.870
Clean26.590.8700.1670.0210.1100.00690.8250.872

As shown in the table, our method maintains reconstruction quality close to the clean setting under moderate perturbations. Stronger perturbations lead to some degradation in rendering and geometric accuracy, but the reconstruction process remains functional. These results demonstrate that our framework can tolerate inaccurate generated supervision under the tested perturbation levels while maintaining effective reconstruction performance.

We further examine the progressive reconstruction process under strong perturbations to assess whether persistent generation errors disrupt subsequent refinement.

Table. Reconstruction progress under strong generation noise.

StagePSNR ↑SSIM ↑LPIPS ↓AbsRel ↓RMSE ↓
Initialization19.460.8160.1980.0500.355
Round 122.990.8490.1790.0320.209
Round 224.060.8540.1790.0290.180
Round 325.470.8620.1770.0260.141
Round 425.990.8660.1770.0230.121
Round 526.130.8690.1740.0230.125

Despite persistent generation perturbations, rendering quality continues to improve across all five reconstruction rounds, accompanied by an overall reduction in depth error. These results show that progressive refinement remains effective under strong generation noise, without interruption of the reconstruction process.

Overall, the experiments demonstrate that GenAnchor maintains reconstruction stability and continues to improve scene quality under the tested generation errors, consistent with the benefits of its observation-based protection and supervision mechanisms.

Reviewer_PZ79OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents GenAnchor, a progressive sparse-view reconstruction framework that uses real observations to guide the application of generative priors. The method combines coverage-aware camera exploration, a tri-state partition that separates grounded, provisional, and uncovered regions, and observation-anchored depth recovery from repaired RGB sequences. Repaired RGB-D views progressively update a 2DGS scene, while grounded regions are excluded from direct synthetic supervision. Experiments evaluate rendering quality, surface geometry, and reconstruction in observed and initially unobserved regions.

Critical strengths
  1. The experimental evaluation covers multiple datasets and input-view settings, with clear improvements in rendering and surface reconstruction over strong baselines.
  2. The appendix provides detailed additional experiments, including separate evaluations of observed and unobserved regions, surface recovery, and progressive coverage. These analyses help assess whether completion gains are accompanied by preservation of observed content.
  3. The paper is clearly written and easy to follow, with a well-organized presentation of the motivation, method, and experiments.
Critical weaknesses
  1. Test-based run selection raises concerns about bias and result stability. Appendix A.7 states that scan5 uses the run with the highest test-view PSNR among three complete reruns. Selecting the best test result can bias both aggregate scores and ablation comparisons. Reporting all three runs, their mean and variability, and a consistent reporting protocol would make the conclusions more trustworthy. The authors should also quantify and explain the differences between runs despite identical seeds and settings, and identify which parts of the pipeline contribute to this variability.
  2. Observation anchoring may preserve geometrically consistent but incorrect estimates. Geometric consistency does not necessarily imply correctness, particularly when verification and depth anchors rely on fixed estimated input depths. Incorrectly accepted regions may therefore retain early reconstruction errors and limit subsequent correction. Although the limitations acknowledge this risk, the paper lacks a targeted robustness analysis showing how the method behaves under erroneous correspondences or depth anchors.
  3. Computational costs are not evaluated. The framework combines repeated video repair, depth recovery, trajectory planning, and scene optimization, but reports neither runtime nor peak GPU memory. Per-scene total time, a module-level breakdown, and peak memory comparisons with baselines on the same hardware are needed to assess the trade-off between reconstruction quality and computational cost.
Questions for authors

Could restoring projected real-input colors in Grounded regions introduce noticeable color or appearance mismatches with the generated completion, especially near their boundaries or on view-dependent surfaces? Please discuss how this is handled and provide representative examples or an analysis of boundary consistency.

5 comments

[W1] Run Variability and Updated Replica Results

We thank the reviewer for pointing out this issue. Selecting a run based on test-view PSNR was inappropriate. We use the mean of all three runs as the scan5 result and update the Replica averages accordingly. We list the results of the three scan5 runs in the table below, together with their mean, standard deviation, and range.

Table. Results of three runs on Replica scan5.

RunPSNR ↑SSIM ↑LPIPS ↓AbsRel ↓RMSE ↓CD ↓F-Score ↑NC ↑
A24.940.8300.1920.0660.1720.01230.6710.793
B24.940.8280.1950.0670.1680.01140.6710.785
C25.140.8330.1970.0710.1760.01220.6590.790
Mean25.010.8300.1950.0680.1720.01190.6670.789
Sample SD0.120.0020.0020.0020.0040.00050.0070.004
Range0.200.0050.0040.0050.0080.00090.0130.008

For each metric, we use the three-run mean for scan5, keep the results of the other seven scenes, and compute the equally weighted average across all eight Replica scenes.

Table. Original and updated Replica averages.

ResultsPSNR ↑SSIM ↑LPIPS ↓AbsRel ↓RMSE ↓CD ↓F-Score ↑NC ↑
Original25.460.8560.1860.0260.1140.00570.8640.876
Updated25.440.8560.1850.0250.1140.00560.8650.876

With this more rigorous reporting procedure, our method maintains its state-of-the-art performance among the compared methods: the rankings of all six main-table metrics remain unchanged. The updated PSNR is 25.44 dB, and the PSNR gain over the preceding ablation setting in Table 2 remains 0.19 dB. We further analyzed the run-to-run variations observed under identical experimental settings and fixed random seeds.

Differences in initial Gaussian counts. The initialized models differ in size, with 6,243,930, 6,244,521, and 6,245,568 initial Gaussians. Thus, divergence is already present before the first RGB diffusion call.

Variability in diffusion repair. The diffusion noise seed is fixed, but rendered RGB, masks, and camera conditions can differ across runs because the Gaussian models used for rendering and planning already differ. These conditioning differences can produce different repaired images even with the same initial noise.

Error propagation through progressive reconstruction. Repaired RGB affects depth recovery, supervision, and Gaussian seeding. The updated scene then determines the next trajectory and generation conditions, allowing early differences to propagate through the five reconstruction rounds.

These observations indicate that run-to-run differences arise before RGB diffusion and can propagate through the progressive reconstruction process, leading to variations in the final results even with fixed random seeds.

We have updated the corresponding results in Tables 1, 2, and A1, as well as the description in Appendix A.7, in the revised version to consistently report the mean of the three scan5 runs. We appreciate the reviewer's careful assessment, which has helped us improve the rigor and transparency of our experimental reporting.

[W2] Sensitivity to Imperfect Input Geometry and Incorrect Anchors

We thank the reviewer for highlighting the need for a targeted robustness analysis. We evaluate how reconstruction evolves and the final depth and surface quality achieved when erroneous estimates pass verification and become depth anchors. We conducted a full reconstruction experiment on Replica scan3, tracking a selected region where errors in the original estimated input depths had been accepted as anchors.

Before the first progressive fitting round, we selected the largest connected region of accepted anchors with relative depth error above 10% in an image from a camera viewpoint chosen in advance. The selected region contains 5,799 pixels, each supported by at least two input views; these erroneous anchor values are retained in the depth supervision. We fixed the region and camera and tracked the reconstructed depth from initialization through all five rounds. The region is selected based on errors in the estimated input depths, whereas the reported metrics are computed from the depths rendered by the reconstructed Gaussian model. Ground-truth depth is used only for this offline analysis and is not involved in reconstruction or anchor verification.

Table. Reconstruction behavior under erroneous depth anchors on scan3.

MeasurementInit.Round 1Round 2Round 3Round 4Round 5
Erroneous-region AbsRel ↓0.108930.108840.109250.109230.109010.10900
Erroneous-region RMSE ↓0.379430.378780.380350.380060.379390.37942
Erroneous-region Bad10 (%) ↓85.8986.9187.0088.4587.1287.39
Valid reconstructed depth pixels (%) ↑84.1799.9799.8699.9899.9399.94

Table. Final depth and surface reconstruction on scan3.

CaseAbsRel ↓RMSE ↓CD ↓F-Score ↑NC ↑
scan30.0210.1100.00670.8340.875

Both tables describe the same run. Bad10 measures the fraction of reconstructed depths in the fixed region with relative error above 10%; the valid-depth percentage is the proportion of pixels with valid reconstructed depth in the full image from the fixed camera, evaluated only where ground-truth depth is available.

As shown in the tables, the local AbsRel remains approximately 0.109 throughout all five rounds, while Bad10 increases slightly from 85.89% to 87.39%. This indicates that the reconstruction process does not substantially correct the errors in the selected region. This behavior is consistent with the use of fixed estimated input depths for geometric verification and anchoring. Depth errors shared across input views can satisfy consistency checks, and accepted values are retained during depth recovery. Subsequent fitting is therefore constrained by geometry that can be internally consistent but inaccurate, allowing a residual bias to persist. Meanwhile, generation and optimization fill missing regions within the same scene, increasing the proportion of pixels with valid depth from 84.17% to 99.94% without adding independent measurements of the true surface. The tracked region remains fully valid throughout, so the persistent error reflects inaccurate reconstructed depths on the same pixels. The final table measures the depth and surface quality achieved under these constraints; the higher local AbsRel (0.10900 versus the scene mean of 0.021) shows that the affected region remains less accurate than the reconstruction on average.

These results show that geometrically consistent but inaccurate depth anchors can remain associated with persistent local reconstruction errors. Although progressive reconstruction increases valid depth coverage, the errors in the selected region remain largely unchanged: AbsRel stays around 0.109, and the fraction of pixels exceeding 10% relative depth error increases slightly from 85.89% to 87.39% after five rounds. This analysis highlights a limitation of relying on fixed estimated input geometry: consistency-based verification does not necessarily guarantee geometric correctness, and subsequent reconstruction may not fully correct errors in accepted anchor regions. This analysis has been added to Appendix A.13 of the revised version.

[W3] Computational Cost, Per-scene Measurements, and Module Runtime

We appreciate the reviewer's comments on computational cost. We report the mean training time and peak GPU memory across the eight Replica scenes, measured on NVIDIA RTX 5090 GPUs. Training time includes input preparation, initialization, diffusion generation and repair, and optimization. Times are in minutes. The results are summarized below.

Table. Computational cost on Replica.

MethodMean training (min)Mean peak GPU memory (GB)
Difix3D+51.9710.12
ArtiFixer105.9024.01
WorldStereo28.8911.12
GuidedVD88.0232.04
G4Splat66.4732.31
Ours67.3417.04

Compared with G4Splat, our method achieves higher mean PSNR with a comparable training time (67.34 vs. 66.47 minutes) while reducing mean peak GPU memory from 32.31 GB to 17.04 GB (47.3%). Our method also requires less training time and peak GPU memory than ArtiFixer and GuidedVD. Overall, these results demonstrate that our method achieves better reconstruction quality with comparable or shorter training time and substantially lower GPU memory usage than these baselines. These efficiency results have been added to Appendix A.10 of the revised version.

To clarify our method's computational cost across scenes and its distribution across processing stages, the following tables report the training time for each Replica scene and the average runtime of each main stage.

Table. Per-scene training time of our method.

Scenescan1scan2scan3scan4scan5scan6scan7scan8Mean
Ours time (min)65.4570.9766.5469.5066.4966.1066.0467.6567.34

Table. Runtime of the main stages of our method.

Main stageMean time (min)
Input preparation, initialization, planning, and condition rendering16.51
RGB diffusion repair and depth recovery9.16
Progressive scene optimization and other pipeline overhead41.67
Complete training67.34

[Q1] Boundary Color Consistency and View-dependent Appearance

We thank the reviewer for raising this concern. To reduce color and appearance inconsistencies between Grounded regions and generated completions, we extend observed Grounded colors to construct a reference image and apply conservative latent replacement throughout denoising. The color extension is used only for the reference and leaves the original Grounded mask unchanged. After each denoising update, latent cells whose pooling regions lie entirely within Grounded are replaced with reference latents at the corresponding noise level, while cells spanning Grounded and editable regions remain free to evolve. Meanwhile, rendered colors in Provisional regions and real reference images provide context, allowing generated content near boundaries to adapt to preserved observations. After decoding, we restore the exact projected real-input colors within the original Grounded mask. For cross-view appearance consistency, camera conditioning specifies each target viewpoint, while real reference images provide appearance guidance during generation. The repaired views are then used to progressively optimize a shared Gaussian scene under real-input supervision, helping maintain coherent surface appearance across different viewpoints. Together, these mechanisms help preserve observed appearance while reducing boundary discontinuities and promoting cross-view consistency.

To further evaluate boundary appearance consistency, we have added before-and-after color restoration examples in Figure A6 of the revised Appendix A.8. The results show continuous structures and natural shading transitions between Grounded regions and generated content, demonstrating that our method effectively maintains boundary appearance consistency.

To evaluate appearance consistency across different viewing directions, we have also added multi-view appearance comparisons of the final reconstruction results in Figure A7. The results show coherent surface colors and appearance across the evaluated test views, without noticeable abrupt color transitions. These examples further support the ability of our framework to maintain consistent appearance across viewpoints while preserving observed content.

The corresponding explanations and figures have been added to Appendix A.8 of the revised version, with the additions highlighted in blue.

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

GenAnchor aims to preserve observed content during generative sparse-view 3D reconstruction. Built on G4Splat initialization and 2DGS, it combines coverage-aware camera exploration, observation-verified tri-state masking, and depth recovery anchored by estimated input geometry. Five view experiments on Replica, ScanNet++, and DL3DV, with additional validation on Mip-NeRF 360, support the complete pipeline’s effectiveness.

Critical strengths
  1. Good writing and organization. The paper presents its motivation, analysis, method, and experiments in a coherent sequence. Detailed derivations. 2.The geometric improvements are supported by complementary metrics and extend to initially unseen regions. On ScanNet++, compared with G4Splat, normalized CD decreases from 0.0103 to 0.0072, F-Score increases from 0.724 to 0.793, and test view depth RMSE decreases from 0.163 m to 0.101 m.
Critical weaknesses
  1. The ablations do not isolate observation anchoring. The OC variant jointly introduces geometry prediction, anchoring, depth supervision, and Gaussian seeding (Table 2; Appendix A.7). Its gains therefore cannot establish the specific contribution of anchoring. Figure 2 also uses different generative backbones across methods.
  2. Test based run selection compromises the Replica evaluation. Appendix A.7 reports selecting one of three scan5 runs by test-view PSNR and using that run in Tables 1, 2, and A1. The resulting selection bias is unquantified and particularly relevant to the modest 0.21 dB gain attributed to OC. This does not invalidate the other datasets, but weakens the reported Replica evidence.
Questions for authors
  1. Please report all three scan5 runs and Replica averages without test-based selection. Do the main and ablation gains persist?
  2. With geometry prediction, repaired images, depth supervision, and seeding fixed, what improvement comes specifically from observation anchoring?
3 comments

[W1/Q2] Ablation Study on Observation Anchoring

We thank the reviewer for raising this concern. We conducted this ablation on Replica by disabling observation anchoring while keeping geometry prediction, repaired images, depth supervision, and Gaussian seed initialization fixed. The table below reports the anchor-disabled result alongside our final results.

Table. Observation anchoring ablation results on Replica.

MethodPSNR ↑SSIM ↑LPIPS ↓AbsRel ↓RMSE ↓CD ↓F-Score ↑NC ↑
w/o observation anchoring25.310.8570.1870.0260.1140.00580.8590.873
Ours25.440.8560.1850.0250.1140.00560.8650.876

Compared with the variant without observation anchoring, our full method reduces normalized CD from 0.0058 to 0.0056, increases F-Score from 0.859 to 0.865, and improves NC from 0.873 to 0.876. PSNR also increases by 0.13 dB, while AbsRel and LPIPS show slight improvements. These results demonstrate the contribution of observation anchoring to improving surface reconstruction quality. This improvement is consistent with its role in constraining depth recovery using geometry verified against real input views, helping the recovered geometry remain aligned with observed structures after RGB repair. We have added this experiment to Appendix A.9 of the revised version.

On revisiting Figure 2, we noted that the original comparison involved different generative models. We have therefore updated Figure 2 so that our method and G4Splat both use the same pretrained WorldStereo model. The corresponding description in Appendix C.2 has also been revised and highlighted in blue.


[W2/Q1] Reporting scan5 Runs and Updated Replica Results

We thank the reviewer for pointing out this issue. Selecting a run based on test-view PSNR was inappropriate. We use the mean of all three runs as the scan5 result and update the Replica averages accordingly. We list the results of the three scan5 runs in the table below, together with their mean, standard deviation, and range.

Table. Results of three runs on Replica scan5.

RunPSNR ↑SSIM ↑LPIPS ↓AbsRel ↓RMSE ↓CD ↓F-Score ↑NC ↑
A24.940.8300.1920.0660.1720.01230.6710.793
B24.940.8280.1950.0670.1680.01140.6710.785
C25.140.8330.1970.0710.1760.01220.6590.790
Mean25.010.8300.1950.0680.1720.01190.6670.789
Sample SD0.120.0020.0020.0020.0040.00050.0070.004
Range0.200.0050.0040.0050.0080.00090.0130.008

We use the three-run mean for scan5, keep the results of the other seven scenes, and compute the average across all eight Replica scenes.

Table. Original and updated Replica averages.

ResultsPSNR ↑SSIM ↑LPIPS ↓AbsRel ↓RMSE ↓CD ↓F-Score ↑NC ↑
Original25.460.8560.1860.0260.1140.00570.8640.876
Updated25.440.8560.1850.0250.1140.00560.8650.876

With this more rigorous reporting procedure, our method still maintains its state-of-the-art performance among the compared methods. The updated PSNR is 25.44 dB, and the PSNR gain over the preceding ablation setting in Table 2 is 0.19 dB.

We have updated the corresponding results in Tables 1, 2, and A1, as well as the description in Appendix A.7, in the revised version. We appreciate the reviewer's careful assessment, which has helped us improve the rigor and transparency of our experimental reporting.

Dear Reviewer SWyT,

Thank you very much for your thoughtful review and valuable suggestions.

In response to your concerns, we have reported the results of three independent runs on Scan5 and provided additional ablation analyses to further validate our method. We sincerely hope that these results and clarifications adequately address your concerns.

As the discussion period is approaching its end, we would greatly appreciate it if you could take a moment to review our responses and share any further thoughts.

Thank you again for your time and consideration.

Best regards,
The Authors

Thank you for the detailed response and the additional experiments. Your rebuttal addressed most of my concerns. I will keep my original rating.

Comments on the paper

1 comment
Area_Chair_ipBvOpenReview ↗

Please share your post-rebuttal review comments

Dear Reviewers,

Thanks for reviewing this paper. Please take a closer look at the feedback from the authors and see if it has addressed your concerns and if you have follow-up questions for the authors. The discussion session will end soon!

Best, AC for paper 58783

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.