acceptodds
Under review as a conference paper at ICLR 2027

SUMI: Symmetry-guided Upsampling Module with Iterative Diffusion for Point Cloud Completion

Abstract

Point cloud completion commonly follows a coarse-to-fine paradigm, where a low-density coarse shape is first predicted and then upsampled to the target resolution. Although recent methods have improved global structure recovery, the fine stage often remains limited by simple upsampling and insufficient interaction with coarse structural features, making local detail reconstruction challenging. We propose SUMI, a diffusion-enhanced refinement module for coarse-to-fine point cloud completion. Unlike prior diffusion-based completion methods that use diffusion as a standalone point generator, SUMI injects noisy geometric features into cross-attention with coarse structural features, enabling reverse denoising to refine local geometry while preserving global consistency. SUMI can also be integrated into existing coarse-to-fine models as a flexible refinement module. Experiments on PCN, ShapeNet-55/34, and MVP demonstrate consistent improvements over strong baselines. SUMI achieves the best overall CD and F1-score on PCN, reduces CD by up to 16.1% on ShapeNet-55, and obtains the best CD across all output densities on MVP.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
computer vision, point cloud, point cloud completion

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_q583OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SUMI, a symmetry-guided diffusion refinement module for point cloud completion. By incorporating iterative diffusion into a coarse-to-fine framework, the method aims to improve local geometric details while preserving global structural consistency. The paper is well-written and presents extensive experiments across multiple benchmarks. However, I have concerns regarding the clarity of the research motivation, the methodological novelty, and whether the relatively limited performance improvements justify the additional computational overhead.

While this paper is well-presented and supported by extensive experimental evaluations, I remain concerned about its methodological novelty and the relatively limited performance gains compared with the additional computational overhead. In particular, the authors should better distinguish SUMI from existing symmetry-guided and diffusion-based point cloud completion methods and provide a more convincing justification for its design. Further clarifying the specific contributions of SUMI to the point cloud completion task would substantially improve the quality of this paper. Given these concerns, I consider even my current score of 2 to be somewhat generous.

Critical strengths
  • The paper is generally well-written and well-organized, with a clear presentation and good overall readability. The figures and illustrations effectively support the explanation of the proposed framework.

  • The proposed method is evaluated on multiple benchmarks, including PCN, ShapeNet-55/34, and MVP. The quantitative comparisons, qualitative visualizations, and ablation studies provide a relatively comprehensive experimental evaluation and demonstrate the effectiveness of the proposed approach to some extent.

Critical weaknesses
  • Unclear research motivation. The motivation of this paper is not sufficiently clear or convincing. Recovering finer local geometric details is a common challenge that has already been extensively investigated by existing point cloud completion methods, such as SeedFormer [1] and CRA-PCN [2]. Therefore, the authors should further clarify the specific limitations of existing approaches and explain why the proposed diffusion-based refinement module is necessary to address them.

  • Concerns about methodological novelty. SUMI can be viewed as an incremental improvement over SymmCompletion [3] through the introduction of a diffusion-based refinement module. Moreover, the methodological differences between the proposed diffusion mechanism and existing diffusion-based point cloud completion methods, particularly 3DDM [4], should be further elaborated. A more detailed discussion is needed to clarify the unique contributions of SUMI beyond existing symmetry-guided and diffusion-based approaches.

  • Concerns about the effectiveness and computational overhead. I have concerns about whether the performance improvements justify the additional computational cost. As reported in Table A6, the largest SUMI configuration requires approximately 9.7% more peak GPU memory than SymmCompletion. However, the average CD in Table 1 improves only slightly from 6.33 to 6.27 (approximately 0.95%) for the main SUMI configuration. Although these results correspond to different SUMI configurations, they raise concerns about the overall trade-off between completion accuracy and computational efficiency. A more comprehensive efficiency analysis would help establish the practical benefits of the proposed method.

[1] SeedFormer: Patch Seeds based Point Cloud Completion with Upsample Transformer. ECCV 2022.

[2] CRA-PCN: Point Cloud Completion with Intra- and Inter-level Cross-ResolutionTransformers. AAAI 2024.

[3] SymmCompletion: High-Fidelity and High-Consistency Point Cloud Completion with Symmetry Guidance. AAAI 2025.

[4] 3DDM:Physically-based Anisotropic 3D Diffusion Model with 3D Gaussian for Point Cloud Completion. AAAI 2026.

Questions for authors

See Critical Weaknesses.

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

The paper presents SUMI, a conditional diffusion refinement module for coarse-to-fine point cloud completion. It inserts noise-conditioned cross-attention between an LSTNet coarse prediction and SGFormer dense upsampling, using coarse, partial, and symmetry features to guide local refinement while preserving global structure. Experiments on PCN, ShapeNet-55/34, and MVP show generally better completion quality than strong baselines, including up to 16.1% CD reduction over SymmCompletion on the hardest ShapeNet-55 setting and consistently lower CD across MVP output densities. However, the PCN improvement over SymmCompletion is marginal, and results on unseen categories are mixed. The authors also conduct ablation studies to examine deterministic versus diffusion refinement, insertion stage, output density, conditioning features, model capacity, and refinement-stage placement.

Critical strengths
  1. The refinement-stage formulation is well motivated. SUMI places denoising after a coarse geometric proposal has already been formed. This is a sensible way to combine global structural guidance with iterative local refinement, and it avoids requiring the diffusion process to infer the entire object from noise. The insertion-stage experiments support this general design choice: fine-stage placement performs much better than introducing diffusion at the initial coarse stage. The comparison with a trained 200-step deterministic refiner is also a useful attempt to separate the effect of diffusion from that of simply performing more refinement iterations.
  2. The evaluation covers several complementary completion settings. The experiments include category-level PCN results, different missingness levels on ShapeNet-55/34, unseen-category evaluation, and multiple output densities on MVP. The improvements on ShapeNet-55 are particularly notable, reaching a reported 16.1% CD reduction in the hardest setting.
Critical weaknesses
  1. The diffusion formulation still has some ambiguity. Eq. (20) assumes point correspondence between the noisy state and the clean estimate, but the paper uses coarse features as queries and supervises the output with permutation-invariant Chamfer Distance. It is not clear how this correspondence is maintained, or how timestep information is supplied to the denoiser.
  2. The gains look modest given the 200-step sampling cost. On PCN, CD-L1 drops from 6.33 to 6.27 against the reproduced baseline, and the gain over the published 6.28 is only 0.01. I would like to see matched runtime numbers and mean/std over multiple seeds before treating this as a clear improvement.
  3. The ablations do not fully isolate where the improvement comes from. Partial and symmetry features are removed together, so the effect of symmetry alone is unclear. Table 5 also shows that the coarse output changes during end-to-end training, making it difficult to separate the benefit of SUMI from changes in the upstream generator.
Questions for authors
  1. How is the point correspondence required by Eq. (20) established? The clean estimate is decoded from coarse queries, while Chamfer Distance only provides permutation-invariant supervision. Please explain why the noisy state and clean estimate can be combined coordinate-wise in the reverse update.
  2. Can you isolate the refinement gain from changes in the coarse generator? Table 5 shows that LSTNet’s coarse CD changes from 9.81 to 9.70 after joint training. A frozen-coarse-generator control would help quantify how much of the final improvement comes from SUMI itself.
  3. Can you report paired mean and standard deviation over multiple training and sampling seeds for SUMI and the reproduced SymmCompletion baseline? Please also provide single-SUMI inference latency, and peak memory under matched settings, including comparisons with the deterministic and feed-forward baselines.
  4. Can you separately ablate partial-feature and symmetry-feature guidance? In particular, what happens when only FsF_s or only FpF_p is removed? The joint ablation changes CD from 6.27 to 6.30, but it does not show how much symmetry guidance contributes on its own.
Reviewer_ZwuiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces SUMI, a diffusion-based intermediate refinement module for coarse-to-fine point cloud completion. The main pipeline uses LSTNet from SymmCompletion to generate a 1,024-point coarse proposal together with partial and symmetry features. SUMI maintains a 1,024-point noisy state, encodes it, and fuses it with the coarse, partial, and symmetry features through attention. A predicted clean point set is used in a 200-step reverse process. The final denoised points are concatenated with the coarse proposal to form a 2,048-point Level-1 output, which is then upsampled by the inherited SGFormer stage to 16,384 points. The full system is trained end to end with Chamfer losses on the coarse, Level-1, and final outputs. The paper reports improvements on PCN, ShapeNet-55/34, and MVP, and also inserts SUMI into AdaPoinTr. Ablations compare single-step refinement, 200-step deterministic recurrence, diffusion refinement, conditioning choices, model capacity, and insertion stage.

Critical strengths
  • SUMI uses a coarse proposal and host features rather than replacing the complete pipeline, making the integration conceptually simple.

  • The paper reports PCN, ShapeNet-55/34, and MVP results, including seen and unseen categories and multiple output densities. The comparison among single-step, 200-step deterministic recurrence, and 200-step diffusion directly tests whether the result follows only from repeated application of the refiner. The insertion-stage experiment is also a useful negative result.

Critical weaknesses
  • The main method uses 200 reverse steps. The appendix reports aggregate evaluation time and peak memory, but it does not report per-shape latency for the baseline, single-step SUMI, deterministic 200-step refinement, and diffusion 200-step refinement on the same hardware. For a refinement module intended to be inserted into existing pipelines, this comparison is essential. A fair compute-matched comparison with a larger deterministic refiner would also be informative.

  • The clean-point posterior mean combines x_t and the predicted clean point set coordinate-wise, and the appendix explicitly states that this assumes coordinate correspondence. However, SUMI is trained with Chamfer distance on the 2,048-point merged output, which is permutation invariant and does not directly enforce correspondence between the 1,024-point state and the predicted clean points. The paper should explain how point identity is maintained.

  • The paper justifies the absence of diffusion baselines by saying prior diffusion methods “often operate on limited ShapeNet subsets” (Sec. 4.1). This is no longer true: PCDreamer (CVPR 2025) and DiffPCN (2025) are diffusion-based completion methods evaluated on PCN with per-category tables and on MVP, and DiffPCN reports the lowest CD and highest F1 on PCN. DiffComplete (NeurIPS 2023) is also absent. Non-diffusion 2024–2025 methods that belong in Tables 1–2 include PointSea (IJCV 2025, PCN CD 6.35) and GeoFormer (ACM MM 2024, PCN CD 6.42). PDR appears only in the MVP table.

  • On ShapeNet-34 Unseen-21, SUMI is worse than SymmCompletion on CD-S and tied on CD-M; on MVP at 2,048 points, F1 decreases despite a small CD improvement. These are not severe failures, but they suggest that the refinement is not uniformly beneficial and should be discussed rather than summarized only as competitive performance.

  • The evaluation uses standard synthetic object-completion benchmarks. A qualitative real-scan test or a controlled noise study would help determine whether stochastic refinement survives sensor artifacts and inaccurate partial observations. This is particularly relevant because the appendix acknowledges that severe real sensor noise is outside the current evidence.

Questions for authors

Refer to my previous comments in Weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.