acceptodds
Under review as a conference paper at ICLR 2027

PRISM: Probabilistic 3D Restoration via Inter-band Spectral Modeling

Abstract

Real-world 3D point clouds are often affected by mixed corruptions, including noise, occlusion, and background interference. Existing methods generally lack explicit physical modeling of corruption, making it difficult to distinguish stochastic disturbances from true geometry or use object-level geometric context to constrain local completion. We therefore reformulate restoration as a “physics-aware spectral generation” problem. Using a fixed 3D discrete wavelet transform, we analyze the frequency-band responses of four corruption operators, express observation changes as operator-induced coefficient deviations, and characterize band-wise differences between stochastic disturbances and true geometry, as well as cross-band relations between the low-frequency skeleton and high-frequency details induced by structural degradation. Building on this analysis, we propose PRISM (Probabilistic 3D Restoration via Inter-band Spectral Modeling), a physics-aware wavelet-based generative framework. Following the Knothe-Rosenblatt (KR) rearrangement theorem, we design an Inter-band Dual-Stream Probabilistic Bridge that decomposes complex joint-distribution inference into cascaded inference of the marginal distribution of the low-frequency skeleton and the conditional distribution of high-frequency details. Experiments show that PRISM provides a physically interpretable inference path and significantly outperforms existing methods on challenging mixed-corruption benchmarks. Code will be released soon.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D point cloud completion, mixed corruption, wavelet representation, generative modeling

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_bTVUOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper provides a reasonable motivation for using wavelets to decompose the process of completing point clouds into coarse features (overall geometric structure) and fine features (including object edges and noise). The two features interact with each other to achieve more accurate completion. Comparative experiments show that this method performs best, and ablation studies also demonstrate the effectiveness of the proposed module. The writing of the paper is clear and easy to read, and the entire process is also clear and easy to reproduce. However, the performance improvement of this paper on the ScanNet dataset is limited. Considering that ScanNet is a dataset composed of real-world data, it is necessary to analyze the reasons for the limited performance improvement. In addition, the ablation experiment for fine->coarse interaction only verified that the weight change is limited to 0.15-0.75, without verifying the case of no fine->coarse interaction (weight=0).

Critical strengths
  1. Well-motivated architecture. Explicit frequency decomposition and coarse<->fine interaction provide an intuitive framework for coordinating global structure and local detail recovery.

  2. Strong synthetic benchmark results. Proposed method improves over baselines on both known-category and unseen-category synthetic shape completion benchmarks.

  3. Informative ablation studies. Component ablations support the effectiveness of cross-stream interaction, joint attention, and persistent observation conditioning.

Critical weaknesses
  1. Limited gains on real scans. Improvements on ScanNet are small, providing limited evidence of robustness to real-world degradation. Further analysis of these modest gains is needed.

  2. Limited evaluation of ambiguity. A incomplete input may correspond to multiple reasonable geometric structures, and using CD alone to measure the accuracy of completion is insufficient.

  3. Limited fine->coarse ablation. The ablation experiment for fine->coarse interaction only verified that the weight change is limited to 0.15-0.75, without verifying the case of no fine->coarse interaction (weight=0).

Questions for authors
  1. Could the authors analyze why the gains on ScanNet are limited and discuss the robustness of the proposed method to real-world degradation?

  2. How does the method handle ambiguity when an incomplete input corresponds to multiple reasonable geometric structures? Could the authors provide evaluation beyond CD and IoU to assess these alternatives?

  3. Could the authors add an ablation with the fine->coarse interaction weight set to 0, while keeping coarse->fine interaction and other components unchanged, to isolate the contribution of fine->coarse interaction?

  4. Could the authors clarify whether PRISM and baselines, particularly BridgeShape, use comparable data augmentation and training settings?

  5. Could the authors provide direct visual comparisons with the strongest baseline, BridgeShape, on the same inputs to demonstrate differences in overall structure and fine details?

5 comments

Thank you for recognizing the architectural motivation and experimental results, and for your constructive suggestions. Below, we address real-scan performance, completion ambiguity, interaction ablations, and comparison settings.

1. ScanNet Gains and Robustness to Real-World Degradations (W1; Q1)

The original manuscript already evaluates naturally occurring mixed degradations. Specifically, §4.1 explicitly states that ScanNet inputs simultaneously contain sensor noise, background clutter, and extreme sparsity, and Table 3 reports the corresponding quantitative results.

Since BridgeShape has not fully released the corresponding ScanNet training and evaluation configurations, we uniformly use zero endpoint noise across all datasets in the main text in the absence of directly reusable parameters. Testing different endpoint-noise levels on ScanNet with a fixed model and three sampling seeds shows that an endpoint-noise level of 0.5 further improves the average CD/IoU, indicating that moderate endpoint perturbations help recovery from real observations.

We will also add the complete endpoint-noise ablation results to the appendix.

ScanNet endpoint noiseCD ↓IoU ↑
06.95153.859
0.56.87354.166

We further conducted experiments with coordinate perturbations, structural interference, missing observations, and field-value noise on the full ShapeNet test set. For each operator, a subtype and a strength level are randomly selected and applied, as specified below.

OperatorSubtypeThree strength levels
G: coordinate perturbationsRotation; scaling; translation0.5°/0.75°/1°; 0.75%/1%/1.25%; 1/2/3 voxels
C: structural interferenceSpherical; wall; floor clutterSphere radius: 1/2/4 voxels; patch width: 3/5/9 voxels
O: missing observationsOcclusion2%/3.5%/5%
N: field-value noiseGaussian noiseσ=0.25/0.30/0.40

The single-operator experiments apply G, C, O, and N separately; the mixed-degradation experiment applies each of the four operators once to every input, with a randomized order. The results are shown below.

ConditionPRISM CD/IoUCD increase / IoU dropPatchComplete CD/IoUCD increase / IoU drop
No additional perturbation3.603/72.951—4.303/65.229—
G: coordinate perturbations4.878/66.8191.275/6.1326.725/57.7402.422/7.489
C: structural interference3.966/70.2490.363/2.7014.755/62.4450.452/2.784
O: missing observations3.961/69.1740.357/3.7764.955/59.4690.652/5.760
N: field-value noise3.765/70.8140.162/2.1377.278/56.9632.975/8.266
Four-factor mixture6.139/58.0642.536/14.88611.876/39.9057.573/25.324

2. Evaluation Metrics and Multiple Plausible Completions (W2; Q2)

We follow the corresponding benchmarks in using CD/IoU and report occupancy F1 in the appendix to evaluate surface distance and occupancy overlap for a single completion. The supplementary evaluation on the full ScanNet test set under the same protocol additionally reports the surface F-score.

MethodScanNet surface F-score@1% ↑
PRISM, default settings24.663
PatchComplete, re-evaluated using the official public weights19.175

The default evaluation uses zero endpoint noise, yielding a deterministic completion for each input; enabling endpoint noise allows the same model to generate multiple results from different endpoints. Thus, although OT-ODE performs deterministic transport given an endpoint, random endpoints still induce a conditional distribution over outputs.

To examine this sampling mechanism, the supplementary experiments on the full ShapeNet and ScanNet test sets generate ten results per input and report both differences between samples and average reconstruction quality.

DatasetEndpoint noiseMean CD ↓Mean IoU ↑Mean pairwise TMD×10⁴
ShapeNet03.60372.9510
ShapeNet0.53.61672.8685.993
ShapeNet13.69771.7098.700
ScanNet06.95153.8590
ScanNet0.56.88154.1395.367
ScanNet17.02252.72212.499

This table shows the controllable effect of endpoint noise on geometric diversity and recovery quality.

3. Ablation Removing High→Low Feedback (W3; Q3)

The ablations in the main text (Table 4) already validate the role of cross-band interaction, while Table 7 examines sensitivity to the feedback coefficient; based on these results, we selected 0.25 as the High→Low coefficient. We further trained under the same settings, setting only the High→Low interaction coefficient to 0. The results are shown below.

High→Low coefficientCD ↓IoU ↑
03.86370.879
0.253.60372.951

The results show that detail feedback helps recover coarse structure. We will add this comparison to §4.3 and Appendix D.1.

4. Comparability of Data Augmentation and Training Settings (Q4)

BridgeShape reports a bridge training budget of 200k steps, using a batch size of 24, AdamW, and a learning rate of 2×10−42\times10^{-4}. PRISM uses the same batch size, optimizer, and learning rate for bridge training on ShapeNet, with a training budget of 100k steps.

Regarding data augmentation, the code released by BridgeShape only targets EPN, on which neither the public implementation of PRISM nor that of BridgeShape enables spatial data augmentation. For ShapeNet training, recent baselines such as BridgeShape and DiffComplete have not released their corresponding data augmentation procedures, whereas the public implementation of PatchComplete explicitly uses random flips, x/z-axis swaps, and ±2-voxel jitter.

We compare data augmentation on ShapeNet in Appendix D.3 (Table 10), with the following results.

ShapeNet augmentation settingCD ↓IoU ↑
No Aug.4.11767.944
Yaw-4 only3.61671.982
PC-Aug3.59972.952

The results show that even when retaining only Yaw-4 and removing PC-Aug, PRISM still achieves 3.616/71.982, outperforming BridgeShape’s 4.06/70.1 in Table 2. This indicates that the gains do not depend entirely on PC-Aug.

5. Visual Comparison with BridgeShape on the Same Inputs (Q5)

Figure 4 and Appendix H of the original manuscript already show structure and detail recovery across datasets and categories. Since BridgeShape has not yet provided the corresponding pretrained weights, a fair visual comparison of the final outputs is difficult.

Thank you for the responses and additional experiments. The fine->coarse weight=0 ablation resolves my concern, and I accept the explanation regarding the unavailable BridgeShape pretrained weights. The following concerns remain:

1. Limited gains on real scans. Adjusting the endpoint noise yields only modest gains on ScanNet, and the response does not explain why this adjustment improves performance. In addition, the main concern is the limited improvement on real-world data, rather than robustness to noise. Additional experiments with predefined synthetic noise do not directly address this concern.

2. Limited evaluation of ambiguity. An incomplete input may admit multiple reasonable completions. My concern is how plausible completions that differ from the ground-truth shape are evaluated (current metrics are insufficient to evaluate this issue). Reporting sample diversity and average reconstruction quality does not directly address this issue.

4. Training comparability. The clarified settings and Yaw-4 results partially address my concern. However, Yaw-4 is still data augmentation, and BridgeShape’s corresponding settings remain unclear, leaving the effects of architecture and augmentation insufficiently separated.

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

The paper addresses 3D shape completion under mixed corruptions. The key idea is to perform completion in a wavelet latent space, using a dual-stream diffusion bridge that recovers the low-frequency structure and the high-frequency details with asymmetric cross-band interaction. The authors model real-world scanning degradation with four physical operators: geometric deformation, external interference (e.g., walls, floors, and other objects), occlusion, and sensor noise.

Critical strengths
  1. The paper is well structured and clearly written, and the authors have committed to releasing the code. The method achieves the best average results across all three benchmarks, with particularly strong improvements on unseen categories in ShapeNet.

  2. The authors break down real-world scan corruption into four types and show how each affects different wavelet components. This helps explain why they use the wavelet domain for 3D shape completion.

Critical weaknesses
  1. The degradation analysis is not well connected to the method or experiments. The four propositions are mostly straightforward properties of the linear wavelet transform. The network does not explicitly use the four corruption operators, and the frequency-band responses are only illustrated schematically, without experimental validation. As a result, the “physics-aware” claim is not well supported.

  2. The novelty of the paper is quite limited, as both wavelet-based latent generation and paired latent bridges have already been explored in previous work (like papers: BridgeShape/ Wavelet Latent Diffusion (WaLa) / Neural Wavelet-domain Diffusion for 3D Shape Generation, Inversion, and Manipulation).

  3. The paper focuses on real-world mixed degradations, but the experiments only use standard shape completion benchmarks. There are no evaluations on mixed-degradation benchmarks, nor comparisons with robust completion methods such as DWCNet.

Questions for authors

The work (Neural Wavelet-domain Diffusion for 3D Shape Generation, Inversion, and Manipulation) already generate coarse wavelet coefficients first and then predict detail coefficients conditioned on them. What is the essential difference between PRISM’s low-to-high conditional design and this coarse-to-detail approach, as well as WaLa’s wavelet latent generation?

4 comments

Thank you for your thoughtful comments on the connection between degradation analysis and our method. Below, we clarify the distinctions from prior work and provide additional experiments on frequency-band responses and mixed degradations.

1. Novelty and Differences from Existing Methods (W2; Q1)

Regarding novelty, the core contribution of PRISM is to formalize mixed degradations as operators on distance fields and analyze their different effects on the two frequency bands and the complementary geometric information retained by each. Based on this analysis, we design a dual-stream architecture to model the recovery of structure and detail separately and exploit their complementary information through bidirectional cross-band interaction. Since high frequencies also contain noise and clutter, we use weaker reverse feedback. This translates the degradation analysis into dual-stream recovery and its asymmetric interaction mechanism.

None of the 3D generation and completion methods with which we compare adopts an interactive dual-stream recovery architecture guided by spectral analysis of mixed degradations. Specifically, Neural Wavelet-domain Diffusion for 3D Shape Generation, Inversion, and Manipulation first generates coarse coefficients through diffusion and then regresses detail coefficients from them. Eq. (9) in PRISM decomposes the observation-conditioned joint posterior, organizing the guidance of high-frequency detail by low-frequency structure. Together with the analysis of cross-band coupling under structural degradation, we further design asymmetric bidirectional interaction that retains High→Low detail feedback, enabling both bands to recover jointly under persistent observation conditioning (§3.3.1, §3.3.3). This design reflects the correspondence between degradation analysis, conditional modeling, and the interaction mechanism, rather than simply following a one-way coarse-to-detail path. WaLa uses compact wavelet encoding to support high-resolution latent diffusion generation; PRISM uses the wavelet representation to characterize degradation-induced biases and organize the recovery of structure and detail accordingly. BridgeShape establishes a latent-space bridge between incomplete and complete shapes; PRISM further explicitly distinguishes the two frequency bands and designs cross-band interaction based on degradation analysis.

These contributions and distinctions from related work are already explained in detail in §3.2–3.3 and Appendix B. To address potential misunderstandings during reading and further clarify the differences from existing wavelet generation and latent-space bridge methods, we will move the relevant discussion from Appendix B to §2.2.

2. Connection Between Degradation Analysis, the Method, and Experiments (W1)

The connection between degradation analysis, the method, and experiments is reflected in the recovery requirements posed by the four types of degradation. To supplement the schematic in Figure 2, we use the controlled degradation settings listed in point 3 (mixed-degradation evaluation) to empirically measure operator-induced frequency-band changes, examining both perturbation magnitude and directional distribution. Relative perturbation is the ratio of perturbation energy to the original signal energy in the corresponding band; the last three columns report each group’s share of the total high-frequency perturbation energy.

Degradation typeLow-frequency relative perturbationHigh-frequency relative perturbationSingle-axis high-pass shareTwo-axis high-pass shareThree-axis high-pass share
G: coordinate perturbations3.17%37.27%54.97%37.93%7.10%
C: structural interference0.27%0.83%81.81%16.29%1.90%
O: missing observations1.44%20.39%79.57%18.96%1.47%
N: Gaussian noise0.35%10.96%43.64%42.41%13.95%
Four-factor mixture5.00%64.07%62.31%31.34%6.35%

The single-axis high-pass group comprises HLL, LHL, and LLH; the two-axis high-pass group comprises HHL, HLH, and LHH; and the three-axis high-pass group is HHH.

For each operator specifically, G perturbs both structure and local geometry in multiple directions, with the two-axis and three-axis high-pass groups together accounting for 45.03%, corresponding to the coordinated recovery of structure and multidirectional detail. For C, additional structures produce directional responses across frequency bands, and the dominant response directions of floor and wall clutter align with their geometric orientations, supporting the combination of structural and boundary information to handle clutter. For O, missing observations alter both local structure and occlusion boundaries, resulting in 79.57% of the high-frequency perturbation being concentrated in the single-axis high-pass group; this corresponds to recovering missing regions using global structural constraints and retained boundary cues. For N, the energy shares of the three single-axis, three two-axis, and one three-axis high-pass subbands are close to 3:3:1, indicating nearly uniform responses across the high-frequency subbands. This differs markedly from the directional boundary responses of O or C and provides empirical evidence for distinguishing structural loss from random perturbations in the frequency-band representation.

These results show that degradations not only affect both frequency bands but also exhibit distinct directional response characteristics. They provide a basis for dual-stream recovery, cross-band information exchange, and detail feedback under structural constraints, and constitute the physical inductive bias referred to as physics-aware in this work. We will add these empirical results to Appendix B.

Building on this, we clarify that analyzing the operators separately does not assume that the degradations are independent. Each increment in Eq. (29) is computed on the state already degraded by the preceding step; for any specified order, the corresponding intermediate states can be defined and a decomposition of the same form established. After Eq. (29) in Appendix B.5, we will further clarify that the order can be chosen arbitrarily and define the corresponding intermediate states.

To examine the wavelet representation itself, we constructed a spatial-domain method with the same parameter count: the mean of each 2×2×22\times2\times2 voxel block and seven of its original voxel values are fed into two separate branches, and the remaining voxel is reconstructed from the mean and the seven voxel values, while retaining the dual-stream network and interaction structure. The final completion results on the full ShapeNet test set are shown below.

RepresentationCD ↓IoU ↑
Spatial domain: mean + 7 voxels3.781868.2502
PRISM: wavelet representation3.603272.9511

In this comparison, the wavelet representation, as the basis for decomposing structure and detail, leads to better final completion performance. We will add the above spatial-domain comparison to §4.3 to further present the role of the representation.

3. Mixed-Degradation Evaluation and DWCNet (W3)

The original manuscript already evaluates naturally occurring mixed degradations. Specifically, §4.1 explicitly states that ScanNet inputs simultaneously contain sensor noise, background clutter, and extreme sparsity, and Table 3 reports the corresponding quantitative results.

We further conducted experiments with coordinate perturbations, structural interference, missing observations, and field-value noise on the full ShapeNet test set. For each operator, a subtype and a strength level are randomly selected and applied, as specified below.

OperatorSubtypeThree strength levels
G: coordinate perturbationsRotation; scaling; translation0.5°/0.75°/1°; 0.75%/1%/1.25%; 1/2/3 voxels
C: structural interferenceSpherical; wall; floor clutterSphere radius: 1/2/4 voxels; patch width: 3/5/9 voxels
O: missing observationsOcclusion2%/3.5%/5%
N: field-value noiseGaussian noiseσ=0.25/0.30/0.40

The single-operator experiments apply G, C, O, and N separately; the mixed-degradation experiment applies each of the four operators once to every input, with a randomized order. The results are shown below.

ConditionPRISM CD/IoUCD increase / IoU dropPatchComplete CD/IoUCD increase / IoU drop
No additional perturbation3.603/72.951—4.303/65.229—
G: coordinate perturbations4.878/66.8191.275/6.1326.725/57.7402.422/7.489
C: structural interference3.966/70.2490.363/2.7014.755/62.4450.452/2.784
O: missing observations3.961/69.1740.357/3.7764.955/59.4690.652/5.760
N: field-value noise3.765/70.8140.162/2.1377.278/56.9632.975/8.266
Four-factor mixture6.139/58.0642.536/14.88611.876/39.9057.573/25.324

Relative to their respective baselines, PRISM exhibits smaller CD increases and IoU drops than PatchComplete under all four individual operators and their random-order combinations. Under the four-factor mixture, the CD increase and IoU drop of PRISM are 2.536 and 14.886, respectively, compared with 7.573 and 25.324 for PatchComplete, demonstrating a consistent robustness advantage under both individual and mixed degradations.

DWCNet/CPCCD uses corrupted point clouds constructed from PCN, with input construction and training and evaluation protocols that differ from those of this work; therefore, direct comparison is not possible. Our purpose in citing DWCNet is to draw on its degradation taxonomy and define the four corresponding operators in the distance-field domain (Appendix A).

Thank you for asking how the degradation analysis informs PRISM’s design. Points 2 and 3 of our response show how occlusion and noise affect wavelet subbands differently and compare PRISM with PatchComplete under mixed degradations. We’d welcome your thoughts when you have a chance.

Reviewer_wYNuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents PRISM, a spectral probabilistic framework for 3D shape completion under mixed corruptions. The method uses a fixed one-level 3D discrete wavelet transform to separate low-frequency structural information from high-frequency geometric details, a Wavelet VQ-VAE to learn complete-shape representations, and an inter-band dual-stream probabilistic bridge for conditional completion. The paper also formulates geometric deformation, structural interference, occlusion, and signal noise as composable corruption operators. PRISM achieves strong results on 3D-EPN and ShapeNet and obtains a small improvement over existing methods on ScanNet. However, several aspects of the claimed novelty, physics-aware formulation, asymmetric interaction design, and robustness evaluation require further clarification and experimental validation.

Critical strengths

The paper combines 3D wavelet decomposition with a dual-stream probabilistic bridge to handle low-frequency structures and high-frequency details. The overall design is reasonable, and the method achieves good results on 3D-EPN and ShapeNet. The authors also provide several ablation studies and efficiency analyses, which help demonstrate the effectiveness of the proposed components.

Critical weaknesses

The novelty of the method is not fully clear, as previous work has already explored joint low- and high-frequency modeling for 3D shape completion. The paper does not provide enough evidence to show why the asymmetric interaction is necessary. Also, the connection between the physical corruption analysis and the actual network is weak, and experiments under mixed corruptions are limited. The improvements on real-world ScanNet data are small, with no multi-seed results to confirm their reliability.

Questions for authors
  1. The methodological novelty should be clarified more carefully. Previous 3D generative and completion studies have explored wavelet and multiscale representations for coarse-to-fine geometric modeling. Therefore, the novelty of PRISM should be clarified beyond its use of 3D wavelet decomposition and joint low/high-frequency modeling. Sections 2.2, 3.1, and 3.3 introduce the spectral representation, Wavelet VQ-VAE, and dual-stream probabilistic bridge, but Table 4 does not isolate their individual contributions, particularly that of the spectral representation. The authors should clarify the respective roles of the spectral representation, dual-stream bridge, and asymmetric interaction, supported by comparisons with spectral and latent-space completion baselines.
  2. The connection between the corruption analysis and the network is unclear. Sections 3.1, 3.2, and Appendix A model four corruption types and analyze their wavelet-domain responses, but the network does not explicitly estimate these corruptions or use their parameters. Thus, the analysis currently serves mainly as theoretical motivation, and its practical contribution remains unclear. Moreover, the assumed Gaussian white-noise model may not reflect real scanning noise. The authors should clarify the role of the operator analysis, validate its relevance to the evaluated data, and moderate the “physics-aware” claim if necessary.
  3. The necessity of the asymmetric cross-band interaction is not fully validated. Section 3.3.3 uses stronger Low-to-High and weaker High-to-Low interactions, but Table 7 only varies the High-to-Low weight with the Low-to-High weight fixed at 1. It does not compare one-way, symmetric, or alternative bidirectional interactions. Moreover, Proposition 2’s aggregate noise-energy analysis does not establish that asymmetric interaction better separates geometric details from noise. The authors should provide these controlled comparisons and justify the selected interaction weights.
  4. The contribution of the wavelet representation is not independently evaluated. The ablations in Table 4 still use wavelet-decomposed features, so they cannot isolate the contribution of the wavelet representation from that of the dual-stream architecture. Appendix D.1 also lacks a matched non-wavelet baseline. The authors should compare PRISM with a spatial-domain or continuous-latent model using comparable capacity and training settings, and progressively evaluate the effects of wavelet decomposition and cross-band interaction.
  5. The robustness evaluation under mixed corruptions is insufficient. Although the paper targets noise, occlusion, interference, and deformation, the current experiments do not evaluate these factors separately or in controlled combinations. ScanNet and the augmentations in Appendix D.3 provide only aggregate evidence. The authors should report results under different corruption types and severity levels, with corruption-specific analysis and, where feasible, comparisons under a matched input representation.
  6. The ScanNet improvement is small and its stability is unclear. Table 3 shows only a modest improvement over BridgeShape, from 6.99 to 6.95 in CD and from 53.5 to 53.9 in IoU, with PRISM underperforming on some categories such as Bag. Since Appendix E reports only fixed random seeds, the stability of these gains cannot be assessed. The authors should provide multi-seed statistics and category-level failure analysis.
  7. The roles of the VQ-VAE and probabilistic bridge require clarification. The bridge uses continuous pre-quantization latents, and VQ lookup is not performed during inference, so the practical benefit of vector quantization remains unclear. A comparison with a continuous Wavelet VAE is needed. In addition, deterministic OT-ODE inference with zero endpoint noise produces a deterministic completion, and the authors should clarify how this setting relates to the claimed probabilistic modeling.
4 comments

Thank you for your careful review and suggestions. We provide the following additional experiments and clarifications on novelty, methodological design, and experimental validation.

1. Novelty (W1; Q1)

Regarding novelty, the core contribution of PRISM is to formalize mixed degradations as operators on distance fields and analyze their different effects on the two frequency bands and the complementary geometric information retained by each. Based on this analysis, we design a dual-stream architecture to model the recovery of structure and detail separately and exploit their complementary information through bidirectional cross-band interaction. Since high frequencies also contain noise and clutter, we use weaker reverse feedback. This translates the degradation analysis into dual-stream recovery and its asymmetric interaction mechanism.

Among existing 3D generation and completion methods, WaLa uses wavelets to construct a compact compressed representation, Neural Wavelet-domain Diffusion first generates coarse coefficients and then regresses detail coefficients from them, and BridgeShape learns bridge transport from partial to complete shapes in a unified latent space. None of these methods adopts an interactive dual-stream recovery architecture guided by spectral analysis of mixed degradations.

These contributions and distinctions from related work are already explained in detail in §3.2–3.3 and Appendix B. To address potential misunderstandings during reading and further clarify the differences from existing wavelet generation and latent-space bridge methods, we will move the relevant discussion from Appendix B to §2.2.

2. Asymmetric Interaction and Wavelet Representation (W2; Q3, Q4, Q1)

Regarding the specific contributions of the wavelet representation, dual-stream architecture, and asymmetric interaction, the ablations in the main text (Table 4) already validate the roles of the dual-stream architecture (Early-Fusion) and cross-band interaction (w/o Cross-Stream Interaction and related ablations). Table 7 examines sensitivity to the feedback coefficient, based on which we selected 0.25 as the High→Low coefficient. To further examine the asymmetric interaction mechanism, we conducted additional training under the same settings, changing only the High→Low interaction coefficient to 0 and 1. The results are shown below.

High→Low coefficientCD ↓IoU ↑
03.86370.879
0.253.60372.951
13.56572.867

The coefficients 0.25 and 1 (symmetric interaction) appear to yield comparable overall performance. However, we further conducted single-step generation (step = 1) experiments, with the following results.

High→Low coefficientCD ↓IoU ↑
0.253.60172.394
13.63772.126

The model with a coefficient of 0.25 is more stable when the number of sampling steps is reduced. From an inverse-problem perspective, coarse structure constrains detail, while high-frequency edges contain both genuine contours and clutter. Weaker High→Low feedback aims to retain local structural cues while limiting noise propagation into the low-frequency stream, providing a mechanistic rationale for asymmetric interaction and for using structural constraints to distinguish geometric details from noise. Multiple updates allow the model to adjust feedback, consistent with the similar results at three steps and the better performance of 0.25 at one step.

To examine the wavelet representation itself, we constructed a spatial-domain method with the same parameter count: the mean of each 2×2×22\times2\times2 voxel block and seven of its original voxel values are fed into two separate branches, and the remaining voxel is reconstructed from the mean and the seven voxel values, while retaining the dual-stream network and interaction structure. The final completion results on the full ShapeNet test set are shown below.

RepresentationCD ↓IoU ↑
Spatial domain: mean + 7 voxels3.781868.2502
PRISM: wavelet representation3.603272.9511

In this comparison, the wavelet representation, as the basis for decomposing structure and detail, leads to better final completion performance. We will add the above interaction-coefficient, single-step generation, and spatial-domain comparisons to §4.3 and Appendix D.1 to further present the respective roles of interaction design and representation.

3. Connection Between Degradation Analysis and the Method (W3; Q2)

This connection is reflected in the recovery requirements posed by the four types of degradation: G changes both global structure and local boundaries; C introduces additional structures, while O removes existing structures, both producing geometric changes across frequency bands and thus requiring recovery to combine structural and detail information. Meanwhile, perturbations introduced by N enter the high-frequency band alongside genuine geometric details, making low-frequency structural constraints necessary for high-frequency feedback. These analyses correspond to dual-stream recovery, cross-band interaction, and weaker reverse feedback, respectively, and constitute the physical inductive bias referred to as physics-aware in this work. We will further clarify these correspondences in §3.2.

Building on this, analyzing the operators separately does not assume that the degradations are independent. Each increment in Eq. (29) is computed on the state already degraded by the preceding step; for any specified order, the corresponding intermediate states can be defined and a decomposition of the same form established. After Eq. (29) in Appendix B.5, we will further clarify that the order can be chosen arbitrarily and define the corresponding intermediate states.

For noise, Gaussian white noise serves as an idealized model for analyzing the frequency-band responses to field-value perturbations, rather than assuming that all real scanning noise follows this distribution. Coupled degradations in real scans are evaluated on ScanNet. The individual G, C, O, and N operator experiments and their random-order combinations below assess actual recovery capability under different degradations, while ScanNet provides validation under real coupled degradations.

4. Mixed Degradations and ScanNet Gains (W4, W5; Q2, Q5, Q6)

The original manuscript already evaluates naturally occurring mixed degradations. Specifically, §4.1 explicitly states that ScanNet inputs simultaneously contain sensor noise, background clutter, and extreme sparsity, and Table 3 reports the corresponding quantitative results.

Since BridgeShape has not fully released the corresponding ScanNet training and evaluation configurations, we uniformly use zero endpoint noise across all datasets in the main text in the absence of directly reusable parameters. Testing different endpoint-noise levels on ScanNet with a fixed model and three sampling seeds shows that an endpoint-noise level of 0.5 further improves the average CD/IoU, indicating that moderate endpoint perturbations help recovery from real observations.

ScanNet endpoint noiseCD ↓IoU ↑
06.95153.859
0.56.87354.166

We further conducted experiments with coordinate perturbations, structural interference, missing observations, and field-value noise on the full ShapeNet test set. For each operator, a subtype and a strength level are randomly selected and applied, as specified below.

OperatorSubtypeThree strength levels
G: coordinate perturbationsRotation; scaling; translation0.5°/0.75°/1°; 0.75%/1%/1.25%; 1/2/3 voxels
C: structural interferenceSpherical; wall; floor clutterSphere radius: 1/2/4 voxels; patch width: 3/5/9 voxels
O: missing observationsOcclusion2%/3.5%/5%
N: field-value noiseGaussian noiseσ=0.25/0.30/0.40

The single-operator experiments apply G, C, O, and N separately; the mixed-degradation experiment applies each of the four operators once to every input, with a randomized order. The results are shown below.

ConditionPRISM CD/IoUCD increase / IoU dropPatchComplete CD/IoUCD increase / IoU drop
No additional perturbation3.603/72.951—4.303/65.229—
G: coordinate perturbations4.878/66.8191.275/6.1326.725/57.7402.422/7.489
C: structural interference3.966/70.2490.363/2.7014.755/62.4450.452/2.784
O: missing observations3.961/69.1740.357/3.7764.955/59.4690.652/5.760
N: field-value noise3.765/70.8140.162/2.1377.278/56.9632.975/8.266
Four-factor mixture6.139/58.0642.536/14.88611.876/39.9057.573/25.324

Relative to their respective baselines, PRISM exhibits smaller CD increases and IoU drops than PatchComplete under all four individual operators and their random-order combinations. Under the four-factor mixture, the CD increase and IoU drop of PRISM are 2.536 and 14.886, respectively, compared with 7.573 and 25.324 for PatchComplete, demonstrating a consistent robustness advantage under both individual and mixed degradations.

To address the training-stability question in Q6, we performed ScanNet transfer training with three training seeds and evaluated each model on the full test set using the default zero endpoint noise.

ScanNet training seedCD ↓IoU ↑
34076.95153.860
34086.92553.903
34096.97853.081
Mean ± standard deviation6.951 ± 0.02753.615 ± 0.463

Across the three training runs, CD remains within 6.925–6.978 and IoU varies between 53.081 and 53.903, showing some variation.

5. Probabilistic Modeling and the Role of VQ (Q7)

The default evaluation uses zero endpoint noise, yielding a deterministic completion for each input; enabling endpoint noise allows the same model to generate multiple results from different endpoints. Thus, although OT-ODE performs deterministic transport given an endpoint, random endpoints still induce a conditional distribution over outputs.

To examine this sampling mechanism, the supplementary experiments on the full ShapeNet and ScanNet test sets generate ten results per input and report both differences between samples and average reconstruction quality.

DatasetEndpoint noiseMean CD ↓Mean IoU ↑Mean pairwise TMD×10⁴
ShapeNet03.60372.9510
ShapeNet0.53.61672.8685.993
ShapeNet13.69771.7098.700
ScanNet06.95153.8590
ScanNet0.56.88154.1395.367
ScanNet17.02252.72212.499

This table shows the controllable effect of endpoint noise on geometric diversity and recovery quality. We will also add the complete endpoint-noise ablation results to the appendix.

The above transport takes place in the continuous latent space before quantization, so the codes predicted by the bridge can be decoded directly. The codebook is used for representation learning in the first stage; performing an additional codebook lookup at inference replaces each predicted vector with its nearest codeword. With the same bridge predictions fixed, the results on the full ShapeNet test set are as follows.

Decoding at inferenceCD ↓IoU ↑
Direct decoding3.597872.9519
Decoding after codebook lookup3.597972.9321

The additional lookup brings no benefit. Furthermore, to examine the role of quantization during training, we separately trained a continuous wavelet AE without quantization and its corresponding completion model. The results on the full ShapeNet test set are shown below.

First-stage representationFinal completion CD ↓Final completion IoU ↑
Continuous wavelet AE3.998070.4367
Current VQ-trained representation3.603272.9511

Thank you for your questions about asymmetric feedback and quantization. Points 2 and 5 show that weaker High→Low feedback improves one-step completion over symmetric feedback, and that training with VQ raises IoU from 70.44% to 72.95%. We’d welcome your thoughts on these ablations when you have a chance.

Comments on the paper

2 comments

Dear Reviewers,

Thank you for your time and effort in reviewing this paper. The authors have posted responses to your reviews. If you have not yet done so, please read their responses and leave a follow-up comment visible to the authors, indicating which concerns have been addressed and which remain unresolved.

Please also update your review and rating if your assessment has changed. Your feedback will help the authors address any remaining questions during the discussion period.

Thank you!

AuthorsOpenReview ↗

General Response

Dear Area Chair and Reviewers,

Thank you for your careful review of PRISM. The central finding of our revised manuscript is that a dual-stream bridge improves 3D shape completion by recovering structure and detail simultaneously and allowing local detail to refine the overall shape. Our theoretical analysis of how corruptions affect different wavelet bands motivates this design, and two new controls provide direct evidence for it. With the same voxel information and network capacity, the wavelet representation outperforms the spatial encoding. With coarse-to-fine guidance held fixed, retaining fine-to-coarse feedback improves completion accuracy. We also tested all four corruptions together on identical inputs and found that PRISM lost fewer IoU points from its own baseline than PatchComplete did.

Degradation analysis leads to a different recovery architecture. A sharp edge in a scan may indicate where a missing surface belongs, or it may come from nearby clutter. To examine how different corruptions affect such geometric evidence, we measured the changes they produce in the wavelet coefficients of the same shape. Occlusion placed 79.6% of its high-frequency perturbation energy in the three single-axis high-pass subbands, compared with 43.6% for Gaussian field noise. This difference motivates preserving directional detail while using the surrounding shape to interpret it. PRISM implements that relationship through strong coarse-to-fine guidance and a weaker return path through which local boundaries can revise the structural estimate. Both streams remain conditioned on the original scan as their estimates evolve.

This two-way correction is where PRISM goes beyond the closest prior approaches. WaLa compresses wavelet coefficients into a common latent grid, leaving open how a corrupted detail band should influence the recovery of global structure. Neural Wavelet-domain Diffusion separates coarse and detail coefficients but predicts them in sequence, so a boundary recovered at the detail stage has no path back to correct the coarse coefficients. BridgeShape addresses completion directly through partial-to-complete transport in a unified latent space, where the direction and strength of structure–detail exchange remain implicit. PRISM maintains a separate, evolving state for each band and explicitly controls their exchange. This architecture lets us test the value of the wavelet representation separately from the value of feeding detail back into structure.

Matched controls isolate the contributions of representation and feedback. To isolate the representation, we gave the same interacting streams two invertible encodings of exactly the same ShapeNet voxel blocks. Both preserved the local mean. With the architecture and parameter count fixed, supplying the remaining information as directional wavelet coefficients in place of individual voxel values raised IoU from 68.25 to 72.95. The early-fusion control in Table 4 also lost accuracy when it retained the wavelet coefficients but merged their latents into a single stream. To isolate feedback, we kept both wavelet streams and coarse-to-fine guidance intact. Disabling only the explicit fine-to-coarse connection reduced IoU from 72.95 to 70.88. Weak and symmetric feedback performed comparably at three sampling steps, with a small advantage for weaker feedback at one step. Across the tested settings, the benefit comes from keeping the coarse estimate open to correction by local detail.

The robustness advantage extends beyond a stronger starting point. On the eight unseen ShapeNet categories, PRISM’s advantage over BridgeShape persisted even when yaw rotations were its only spatial augmentation, with mean IoU of 71.98 versus BridgeShape’s published 70.1. On real ScanNet objects, three transfer runs with different training seeds averaged 6.95 CD and 53.62 IoU, close to BridgeShape’s published 6.99 and 53.5. These benchmarks measure accuracy on the observations they contain. To determine how sensitive a method is to corruption, we compared its performance before and after degrading the same inputs.

We therefore evaluated PRISM and PatchComplete on identical observations across the full ShapeNet test set, first applying deformation, clutter, occlusion, and field noise individually. For every corruption, PRISM had a smaller increase in Chamfer distance and a smaller loss in IoU relative to its original performance. The advantage persisted when all four corruptions were applied to every input in random order.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.