{"id":"78c4eb84-3d14-4324-98f7-6ee51082fbdb","arxiv_id":"2608.08115","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"SUMI refines coarse 3D point cloud predictions by injecting noisy geometric features into cross-attention, and reports state-of-the-art Chamfer distance on PCN, ShapeNet-55/34, and MVP.","lead":"The paper proposes SUMI, a module that adds diffusion-style denoising to the fine stage of point cloud completion models, refining local geometry while keeping the global shape. It reports the best scores on three standard 3D point cloud benchmarks, though the gains over the closest baseline are small on PCN.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reverse-sampling update used at inference is never defined, so the claimed diffusion-based refinement and the T=200 gains in Fig. 7 are not reproducible or attributable to a DDPM reverse process.","rationale":"The reader's weakest assumption is exactly the point on which the paper is least secure: Section 3.3.1 invokes the DDPM reverse process but never defines it, and the supplementary material contains only the forward process and the CD training loss. This is not a disagreement with an existing consensus; it is an internal omission that blocks reproduction and prevents attribution of the Fig. 7 T=200 gain to iterative denoising. No code or weights are provided, and the ablation figures report no numeric values or error bars, so the empirical claim rests entirely on a procedure that is not written down. I considered whether the more serious issue is the near-zero performance gap on PCN (Ours CD 6.27 vs SymmCompletion 6.33) or the lack of error bars, but those would only weaken the strength of the claim, whereas an undefined inference rule makes the central mechanism untestable. The appropriate reader verdict remains CONDITIONAL: the authors should specify the reverse update, release code, and report variance; until then the strongest claim should not be taken at face value. My read therefore leaves the reader's verdict unchanged.","tokens_in":13182,"tokens_out":5610,"duration_ms":54455,"concrete_test":"Obtain from the authors the exact reverse-sampling update used at inference (or released code). Then, with the same trained SUMI checkpoint, compare two sampling procedures on the MVP T=200 configuration: (a) the procedure described in the paper, reconstructed as literally as possible, and (b) a proper DDPM reverse step for an x0-predicting network, e.g. xt−1 = sqrt(αt)−1 (xt − βt/sqrt(1−ᾱt) εθ(xt,t)) + σt z, where εθ is derived from the predicted x0. If procedure (a) is not a valid reverse update, or if neither procedure reproduces the reported CD/F1 values, the claim that SUMI performs diffusion-based reverse denoising fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim identifies SUMI as a diffusion-enhanced refinement module that performs reverse denoising (§1, §3.3.1). Training is specified only through the forward noising process q(xt|xt−1) = N(xt; sqrt(1−βt) xt−1, βt I) (Eq. 1) and an x0-prediction loss Ldiff = L_CD(xf, x0) (Eq. 2). No reverse posterior q(xt−1|xt,x0), no noise-prediction parameterization, and no sampling update are given. The main text says 'the predicted clean point cloud is used as the denoised estimate at each step, following the DDPM formulation summarized in the supplementary material'; however, the supplementary material writes only the forward equations and the same CD loss, and never states how xt−1 is formed from xt and xf. If the actual implementation sets xt−1 = xf, or uses some other ad hoc rule, the T-step procedure is not a DDPM reverse transition, and the T=200 vs T=0 ablation in Fig. 7 measures repeated feed-forward refinement, not diffusion denoising. Since the reverse transition is the named mechanism behind the reported improvements, the absence of this rule makes the headline results neither reproducible nor attributable to diffusion.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":13422,"tokens_out":5899,"duration_ms":54956,"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":[{"comment":"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.","section":"§3.3.1 and Supplementary §2"},{"comment":"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.","section":"§3.3.1–3.3.2"},{"comment":"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.","section":"Tables 1–3 and §4.2–4.4"}],"minor_comments":[{"comment":"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.","section":"Eq. (5)"},{"comment":"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.","section":"Tables 2 and 3"},{"comment":"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.","section":"§3.4"},{"comment":"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.","section":"§2.3"},{"comment":"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.","section":"§4.1"},{"comment":"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.","section":"Supplementary Table A2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript builds closely on SymmCompletion [8], using its coarse LSTNet module and fusion-block design; the claimed novelty is confined to the noise-conditioned cross-attention and the diffusion-inspired iterative refinement. Given that, I would ask the editor to require a clear statement of the inference-time update and, ideally, code release before considering publication. The small margins over the baseline (e.g., 6.27 vs 6.33 CD on PCN) mean that the absence of variance reporting is not a minor issue here."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","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.\n\nThe 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.\n\nMinor 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.\n\nThe 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.\n\nMy 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.","headline":"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.","tokens_in":13961,"tokens_out":1899,"would_cite":false,"duration_ms":21272,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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…","keywords":["point cloud completion","diffusion model","cross-attention","coarse-to-fine","refinement module","Chamfer Distance","reverse denoising"],"falsifier":"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.","tokens_in":12931,"feed_emoji":"🧩","tokens_out":9383,"duration_ms":83803,"temperature":0.7,"pith_summary":"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.","feed_headline":"Diffusion inside attention cuts point cloud completion error","feed_subtitle":"Lightweight refinement module improves local detail while preserving global shape on PCN, ShapeNet-55, and MVP.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Provides the coarse generation module and the keypoint and symmetry features that SUMI conditions on, and serves as the strongest baseline in the PCN tables.","marker":"[8]"},{"why":"Supplies the DDPM forward diffusion formulation and the training and sampling paradigm SUMI follows.","marker":"[43]"},{"why":"The transformer-based coarse-to-fine approach SUMI builds on, and the model used for the plug-in integration experiment with AdaPoinTr.","marker":"[5]"},{"why":"The prior conditional diffusion-refinement baseline for point cloud completion, contrasted with SUMI's design of diffusion as a refinement signal rather than a standalone generator.","marker":"[67]"},{"why":"The primary dataset of partial-complete pairs on which the reported PCN CD and F1 numbers are measured.","marker":"[69]"},{"why":"The dataset for the up-to-16.1% CD reduction claim and the generalization evaluation on 21 unseen categories.","marker":"[70]"},{"why":"The multi-view dataset where SUMI claims the best CD at all four output densities.","marker":"[71]"},{"why":"A strong MVP baseline that anchors the comparison table at multiple densities.","marker":"[77]"}],"fun_headline_variants":["Diffusion powers attention to refine point cloud details","Noise-conditioned cross-attention improves point cloud completion","SUMI injects noisy geometry into cross-attention for refinement","Refine point clouds by denoising via attention","Diffusion-enhanced attention for finer 3D point clouds"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion powers attention to refine point cloud details","Noise-conditioned cross-attention improves point cloud completion","SUMI injects noisy geometry into cross-attention for refinement","Refine point clouds by denoising via attention","Diffusion-enhanced attention for finer 3D point clouds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000691,"raw_usage":{"total_tokens":3124,"prompt_tokens":938,"completion_tokens":2186,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":2107}},"tokens_in":554,"tokens_out":2186,"duration_ms":16638,"temperature":1.0,"reasoning_tokens":2107,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:23:18.003126+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"Pointr: Diverse point cloud completion with geometry-aware transformers","cited_arxiv_id":null,"evidence_quote":"The transformer-based coarse-to-fine approach SUMI builds on, and the model used for the plug-in integration experiment with AdaPoinTr."},{"cited_title":"Pcn: Point completion network","cited_arxiv_id":null,"evidence_quote":"The primary dataset of partial-complete pairs on which the reported PCN CD and F1 numbers are measured."},{"cited_title":"3d shapenets: A deep representation for volumetric shapes","cited_arxiv_id":null,"evidence_quote":"The dataset for the up-to-16.1% CD reduction claim and the generalization evaluation on 21 unseen categories."},{"cited_title":"Multi-View Partial (MVP) Point Cloud Challenge 2021 on Completion and Registration: Methods and Results","cited_arxiv_id":"2112.12053","evidence_quote":"The multi-view dataset where SUMI claims the best CD at all four output densities."},{"cited_title":"Snowflakenet: Point cloud completion by snowflake point deconvolution with skip-transformer","cited_arxiv_id":null,"evidence_quote":"A strong MVP baseline that anchors the comparison table at multiple densities."}],"review_version":1}