SUMI: Scalable Unified Model for 3D Point Cloud Inference

Kanchana Thilakarathna; Yanlong Li

SUMI: Scalable Unified Model for 3D Point Cloud Inference

T0 review · 3 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper claims that SUMI, a diffusion-enhanced refinement module that injects noisy geometric features into cross-attention with coarse structural features, can be attached to coarse-to-fine completion models to improve local detail…

desk verdict SUMI is a plausible incremental completion method with a genuinely new cross-attention twist, but the paper never defines its inference-time reverse diffusion step, so the central 'diffusion-enhanced' claim is not yet reproducible or attributable. read the letter →

arxiv 2608.08115 v2 pith:QGLOXQVR submitted 2026-08-08 cs.CV

Yanlong Li , Kanchana Thilakarathna This is my paper · ORCID

classification cs.CV

keywords pointcloudcompletiondiffusionmodelcross-attentioncoarse-to-finerefinementmoduleChamferDistancereversedenoising

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Point cloud completion usually works in a coarse-to-fine fashion: predict a sparse overall shape, then upsample to full density, and the fine stage is often the weak link. SUMI proposes to make the fine stage a diffusion-enhanced refinement module rather than a simple upsampler. The module injects noisy geometric features into cross-attention keys and values, with coarse structural features as queries, so that iterative reverse denoising sharpens local detail while the coarse prediction keeps global structure intact. The paper reports that this design achieves the best overall Chamfer Distance and F1-score on PCN, reduces Chamfer Distance by up to 16.1% on ShapeNet-55, and obtains the best CD at every output density on MVP. It also shows that the module can be dropped into an existing coarse-to-fine model such as AdaPoinTr and improve it without major architectural changes.

What carries the argument

The load-bearing object is the noise-conditioned cross-attention fusion block inside the SUMI module. In each Fusion Block, the feature representation of the coarse prediction acts as the Query and the feature representation of the noisy point cloud acts as the Key and Value; a self-attention layer then enhances the fused features, and separate branches fuse in keypoint features and symmetry features before a final fusion block and a fully connected layer map to the upsampled point cloud. This mechanism lets stochastic perturbations influence how coarse structure and noisy geometry interact, which is what turns the diffusion process into a refinement signal rather than a coordinate generator. The second load-bearing piece is the three-layer coarse-to-fine pipeline that places SUMI at the first fine-generation stage and leaves a lightweight upsampler to reach the final density.

What would settle it

Run the same SUMI architecture on PCN with the mathematically correct DDPM posterior update (adding scheduled noise back at each reverse step) instead of the paper's clean-prediction substitution; if CD does not improve over the T=0 single-step baseline, the reported gains are not produced by a real reverse diffusion process.

Watch

Extended reading notes

Core claim

The central discovery is that diffusion does not need to be a standalone generator to help point cloud completion; it can serve as a fine-stage refinement signal inside feature interaction. SUMI encodes the coarse prediction and the noisy point cloud separately, then fuses them with cross-attention in which the coarse features are the query and the noisy features are the key and value, followed by self-attention and further fusion with keypoint and symmetry features. Because the key and value come from noise-perturbed geometry, the attention weights are implicitly conditioned on noise, and the network learns to predict the clean point cloud from the noisy observation. At inference, the predicted clean point cloud is used as the denoised estimate at each of T steps, with T=200 giving the best result. The paper claims this yields the best overall $\ell^1$ CD and F1-score on PCN, a CD reduction of up to 16.1% on ShapeNet-55, and the best CD across all four output densities on MVP, while also improving AdaPoinTr when integrated into it.

Load-bearing premise

The central claim depends on the assumption that iteratively replacing the noisy input with the network's predicted clean point cloud across T steps is a valid reverse diffusion process, an update rule the paper never states.

Editorial extensions

If this is right

Existing coarse-to-fine completion models can adopt SUMI as a plug-in refinement stage and expect lower CD and higher F1 without retraining the whole architecture from scratch, as demonstrated with AdaPoinTr.
Using diffusion at the fine stage rather than as a generator preserves global consistency and yields better local detail, with the gains attributed to iterative denoising because T=200 outperforms T=0 with the same architecture.
Applying SUMI at the coarse stage degrades performance; the refinement belongs after coarse prediction, so future designs should target the fine-generation stage for diffusion-based refinement.
Increasing inference timesteps improves quality up to T=200 and degrades at T=300, so there is a sweet spot balancing reconstruction quality and inference cost.
The improved quality comes with extra inference cost from T steps of denoising, which the paper mitigates by applying SUMI only at the first fine stage and using a lightweight upsampler for the final density.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper never writes the reverse update rule; it only states the forward process and the training loss. If the iterated clean-prediction substitution is not a proper DDPM posterior step, some of the reported gains could come from the fusion architecture or from timestep conditioning rather than from diffusion itself.
The same noise-conditioned cross-attention design could transfer to other dense 3D tasks such as depth completion or scene completion, where a coarse prediction and a noisy observation must be fused in feature space; that would be a testable extension the paper does not explore.
Because SUMI conditions on keypoint and symmetry features from the coarse model, its gains may depend on the coarse model's symmetry prior; on categories with weak or no symmetry, the refinement may add less.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The thing to know about SUMI is that the architecture is more interesting than the results. Injecting noisy point features as keys and values into cross-attention with coarse features as queries is a real departure from PDR-style point-level diffusion refinement. The ablations on timestep, density, and insertion stage are sensible, and the T=0 versus T=200 comparison is the right kind of control: it suggests the iterative refinement is doing something beyond the added parameters. On the empirical side, the gains over SymmCompletion and AdaPoinTr are consistent on three benchmarks, which is decent evidence that the module helps.

The soft spot is exactly where the stress-test note lands. The paper says at inference it uses the predicted clean point cloud as the denoised estimate at each step, following the DDPM formulation in the supplementary. But the supplementary only gives the forward noising process and the CD training loss; there is no reverse posterior, no x0-to-x_{t-1} update rule, nothing. If the implementation literally sets x_{t-1} = x_f, or uses some other ad hoc rule, then the T-step loop is not a DDPM reverse transition. That matters because the paper’s headline claim is that diffusion-based reverse denoising is what improves local geometry. Without the update rule, the T=200 gains in Fig. 7 could just be repeated feed-forward refinement through the same network. That is a load-bearing gap, not a cosmetic one. It is also easily fixed: write down the exact update, or drop the diffusion label and call it iterative refinement.

Minor issues: no code or weights, no error bars on the tables, and only one external baseline (AdaPoinTr) used for the plug-in demonstration. Those are addressable.

The citation pattern is clean. Using SymmCompletion as the coarse module is transparent, and the related work is fair to PDR and other diffusion-completion methods.

My recommendation: a serious referee should see this. The idea is plausible, the empirical signal is real, and the missing sampling step is a fixable specification problem rather than a fatal flaw. I would ask for the exact reverse update, code release, and variance reporting before accepting. If the authors cannot supply a proper reverse transition, the paper should be reframed as iterative feature-refinement, not diffusion.

Referee Report

3 major / 6 minor

Summary. The manuscript proposes SUMI, a diffusion-enhanced refinement module for coarse-to-fine point cloud completion. The module takes a coarse prediction and a noisy point cloud, encodes them separately, and fuses them through cross-attention in which coarse features serve as queries and noisy features as keys/values; keypoint and symmetry features are then incorporated via additional fusion blocks. Training uses a Chamfer-distance loss to predict the clean fine point cloud from a noisy observation, jointly with coarse and upsampling losses. At inference the authors state that reverse denoising is performed for T steps, with the predicted clean point cloud used as the denoised estimate at each step. Experiments on PCN, ShapeNet-55/34, and MVP report improved CD/F1 over baselines, and ablations study output density, insertion stage, and diffusion timestep T.

Significance. If the reported results are reproducible, SUMI is a practically useful plug-in refinement stage: it improves a strong coarse-to-fine baseline (SymmCompletion) on three benchmarks and can be inserted into AdaPoinTr with minimal architectural changes. The paper also contributes a useful ablation study on the diffusion timestep and the insertion stage, and it evaluates on standardized benchmarks with broad category coverage. However, the lack of code, the absence of error bars, and the under-specification of the inference-time reverse process currently leave the central 'diffusion' attribution unsupported.

major comments (3)

[§3.3.1 and Supplementary §2] The inference-time reverse process is never defined. The main text states that 'the predicted clean point cloud is used as the denoised estimate at each step, following the DDPM formulation summarized in the supplementary material,' but the supplementary only gives the forward kernel q(xt|xt−1), the marginal q(xt|x0), and the loss Ldiff; no reverse posterior q(xt−1|xt,x0), no sampling update, and no initialization for the reverse chain are provided. Consequently, the procedure used to obtain the T=200 results in Fig. 7 cannot be reproduced, and the gains cannot be attributed to a DDPM reverse transition rather than to repeated feed-forward application of the same network. Please provide the exact update rule and the initialization of the reverse chain, and either justify that it is a proper DDPM posterior or revise the diffusion-based claims.
[§3.3.1–3.3.2] The network is not explicitly conditioned on the diffusion timestep t. In standard DDPMs, the model receives t (e.g., through a sinusoidal time embedding) so that one network can handle different noise levels; here the inputs are the noisy point cloud xt and the coarse/symmetry features, with no mention of any t embedding or t-dependent feature. Because the cross-attention is described only as Fc (query) against F_t^x (key/value), the noise level is available only implicitly through the statistics of xt. Without explicit time conditioning, iterative application of the same network for T steps is not a discretization of a known diffusion reverse process, and the meaning of 'timestep T' in Fig. 7 is unclear. Please specify the time-conditioning mechanism or clarify how the model distinguishes different noise levels.
[Tables 1–3 and §4.2–4.4] The empirical claims are reported without error bars or significance tests, and the margins over the strongest baseline are small in several cases (PCN overall CD 6.27 vs 6.33 for SymmCompletion; MVP 2048 CD 4.85 vs 4.89 with F1 0.53 vs 0.54). The 'best' claims in the abstract and in §4.2–4.4 therefore rest on single-run numbers. Please report means and standard deviations over at least three training runs, or make the code and exact evaluation protocol available so that the results can be independently verified.

minor comments (6)

[Eq. (5)] Equation (5) uses the symbol '·' for the operation that is afterwards defined as concatenation; as written, the expression resembles a product of two attention outputs. Please use a dedicated concatenation symbol (e.g., ⊕) or brackets.
[Tables 2 and 3] Several numeric entries in Tables 2 and 3 lack spacing (e.g., '4.890.54', '2.140.85', '0.390.70'), which makes the tables difficult to read; please fix the formatting.
[§3.4] Section 3.4 states that the SUMI input density is 2048 for PCN and is then 'further improved to the target resolution,' but it does not state the output density of the SUMI module itself or describe the lightweight upsampling module; please clarify this processing chain.
[§2.3] In Section 2.3, references [58, 72, 65, 66] are grouped as prior diffusion methods operating on limited ShapeNet subsets, but [72] is a paper on semantic-collaborative diffusion for anomaly detection and does not appear to be a point cloud completion method; please re-check the citation.
[§4.1] The F1-score threshold is not defined in Section 4.1; please specify it (e.g., 1% of the bounding box diameter) so that the reported F1 numbers can be compared with other papers.
[Supplementary Table A2] In the supplementary Table A2, the evaluation time for MVP is 550 minutes per epoch, which seems disproportionately large relative to the other datasets; please clarify whether this includes the full test set and the T-step reverse process.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: SUMI's CD/F1 gains are empirical benchmark results trained end-to-end on Chamfer Distance and evaluated on held-out data; the undefined reverse-sampling update is a reproducibility gap, not a circular reduction.

full rationale

The claimed derivation chain is an empirical coarse-to-fine completion system. The coarse module is imported from external work (SymmCompletion [8]), and the SUMI refinement module is trained end-to-end with the Chamfer Distance objective of Eq. 2 (Ldiff = LCD(xf, x0)) and the composite loss of Eq. 6. The headline numbers (PCN CD/F1, ShapeNet-55 CD reductions, MVP CD at all densities) are measured on standard held-out test splits against external baselines. No fitted constant, fitted scale, or tuned coefficient is relabeled as a prediction; the metric used for evaluation is also the training loss, but that is standard supervised learning, not a circularity. The Fig. 7 ablation compares T=0 and T>0 under the same refinement architecture, so the reported gain is an empirical result rather than a tautology. I also checked for self-citation load-bearing steps: the paper does not rely on the present authors' prior work; the backbone and fusion blocks are attributed to external papers [8, 68], and the DDPM forward process is standard [43]. No uniqueness theorem is imported from the authors. One genuine omission is flagged: Section 3.3.1 states that 'the predicted clean point cloud is used as the denoised estimate at each step, following the DDPM formulation summarized in the supplementary material,' but neither the main text nor Supplementary Section 2 gives the reverse transition xt-1 = f(xt, xf) or the posterior q(xt-1|xt, x0). This is a missing derivation and a reproducibility risk for the diffusion attribution, but it is not an equation reducible to its inputs by construction and therefore does not constitute circularity under the stated criteria.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central empirical claim rests on standard supervised learning with hand-chosen hyperparameters. No new physical entities are introduced. The most fragile axiom is the ad hoc assumption that the training objective, predicting clean x0 with CD loss, supports a valid reverse-diffusion loop at inference, because the paper never specifies the reverse update.

free parameters (5)

Diffusion timesteps T = T=200
Chosen via ablation on MVP in Fig. 7; T=300 degrades performance.
SUMI output density = 2048 points
Chosen via ablation on PCN in Fig. 5; direct generation of 16384 points degrades performance.
SUMI insertion stage = Level-1 fine stage
Chosen via ablation in Fig. 6; coarse-stage insertion degrades performance.
Number of fusion blocks = 4
Architectural choice stated in Section 3.3.2; not ablated.
Training epochs = 350 (PCN/ShapeNet-55/34), 40 (MVP)
Supplementary Table A2; results may depend on these budgets.

assumptions (5)

standard math DDPM forward diffusion is a valid noise model for point clouds
Used in Eq. 1 and Supplementary Eq. 2; standard result from Ho et al. [43].
domain assumption The coarse module output Pc preserves sufficient global structure to condition fine refinement
Section 3.2 adopts LSTNet [8] as the coarse generator; SUMI relies on Pc for global consistency.
domain assumption Keypoint features Fk and symmetry features Fs are informative conditional signals
Section 3.3.2 fuses Fk and Fs as conditioning; their usefulness is inherited from [8], not re-established here.
ad hoc to paper Predicting clean x0 with CD loss yields a valid reverse-diffusion denoiser at inference
Section 3.3.1 and Supplementary Section 2 specify this training objective but never derive the inference update rule.
domain assumption CD and F1 on PCN/ShapeNet-55/34/MVP are accepted measures of completion quality
Section 4.1 adopts these metrics following prior work; all conclusions depend on their validity.

how reviews work

0 comments

Cite this review

Pith. "Pith review of SUMI: Scalable Unified Model for 3D Point Cloud Inference." pith.science (2026). https://pith.science/paper/QGLOXQVR

@misc{pith2026260808115,
  author       = {Pith},
  title        = {Pith review of: SUMI: Scalable Unified Model for 3D Point Cloud Inference},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QGLOXQVR}},
  note         = {Machine review of arXiv:2608.08115}
}

read the original 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.

Figures

Figures reproduced from arXiv: 2608.08115 by the authors.

**Figure 2.** Level-1 fine prediction comparison. From left to right: AdaPoinTr, SymmCompletion, Ours, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Detailed architecture of the SUMI module. The coarse prediction and noisy point cloud are [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Qualitative comparison between our method and prior works. Our method not only [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 5.** Figure 5: Ablation study on output density for SUMI Module [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Continue with ORCID to comment.

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.