{"id":"d2259680-1898-4972-9390-5c676a4705eb","arxiv_id":"2501.15286","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A flow-matching model with Earth Mover's Distance pre-alignment upsamples point clouds in five sampling steps, with state-of-the-art Chamfer distance scores on PUGAN and PU1K.","lead":"The authors train a generative model to turn low-resolution 3D point clouds into high-resolution ones by learning a direct path between them, instead of starting from random noise. Their method adds a matching step during training and reports better quality with far fewer sampling steps than diffusion-based alternatives.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The sampling update in Algorithm 2 contradicts the Euler step in Eq. 4, leaving the reported 5-step results tied to an unspecified integration scheme.","rationale":"In good faith, I read the paper's central claim as: a flow-matching model trained with midpoint interpolation and EMD pre-alignment can upsample point clouds at state-of-the-art quality with very few sampling steps (5 vs 30). For this claim to hold, the training objective (Eq. 6) and the inference-time ODE solver must be consistent, because the learned velocity field ν_θ is defined by the former and consumed by the latter. The paper provides two conflicting definitions of the inference update: Eq. (4) is the standard Euler step for dx/dt = ν_θ, while Algorithm 2's update adds a contraction toward zero that does not solve that ODE. The manuscript never clarifies which one was implemented, and no code is released. Consequently, the reported quantitative results, including the headline 5-step efficiency, cannot be reproduced or even uniquely attributed to a well-defined algorithm. I considered the reader's weakest assumption (EMD pre-alignment training/inference shift) and found it less decisive: flow matching with any coupling between source and target samples still yields a marginal vector field that transports the source distribution to the target distribution, so the absence of alignment at inference is not by itself a distribution-shift failure as long as the source marginal matches. The sampling inconsistency, by contrast, breaks the direct link between the trained objective and the inference procedure. This is not an attack on the authors' integrity; it is a concrete, addressable gap in the method definition. If the code or a clarification shows that Eq. (4) was used and Algorithm 2 is a typo, the concern would be resolved, and the method could return to the CONDITIONAL status the reader assigned. Until then, the central claim is unverified as specified.","tokens_in":11960,"tokens_out":11021,"duration_ms":104072,"concrete_test":"Obtain the authors' code (or, failing that, re-implement both updates) and run 4× upsampling on the PUGAN test set with (i) the standard Euler step of Eq. (4) using uniform step sizes and (ii) Algorithm 2's update, keeping all other settings identical. Compare CD, HD, and P2F at 5 sampling steps. If the two updates differ by more than ~2% relative in any metric, the paper's reported sampling is ambiguous and the efficiency claim is not uniquely defined. Also verify which update reproduces the values in Table 1; this will settle whether the central results are tied to an incorrect integration scheme.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The most load-bearing concern is the internally inconsistent sampling update. Equation (4) defines the Euler step for the learned ODE as x_{t+1} = x_t + (δ_{t+1} − δ_t) ν_θ(x_t, t), which is the standard integration of dx/dt = ν_θ. Algorithm 2 instead states x_{t+1} = (1 − δ_t) x_t + δ_t ν_θ(x_t, t). These updates differ by the extra term −δ_t x_t; the latter is not an Euler step for the trained flow and corresponds to a different discrete process. The paper does not state which update produced Tables 1–4 or the 5-step runtime in Table 3. If Algorithm 2 was used, the reported quality and efficiency results are not grounded in the flow-matching objective of Eq. (6), and the central claim of 'directly learns the optimal transport' is unsupported. If Eq. (4) was used, the pseudocode is erroneous and the method is under-specified without code. Either way, the central claim is not reproducible from the manuscript as written. This is a concrete, correctness-level gap that precedes the secondary question of EMD-alignment generalization at inference.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PUFM, a flow-matching model for point cloud upsampling that maps a sparse point cloud directly to a dense point cloud. The method first densifies the sparse input via midpoint interpolation, pre-aligns the interpolated points to the dense ground truth using Earth Mover's Distance during training, and then learns a conditional velocity field with a flow-matching objective. At inference, the learned ODE is integrated from the interpolated sparse input without alignment. Experiments on PUGAN and PU1K report state-of-the-art Chamfer Distance, Hausdorff Distance, and Point-to-Surface metrics at 4x and 16x upsampling, with only 5 sampling steps, and additional experiments on noisy inputs, arbitrary upsampling factors, ScanNet, and KITTI are presented.","tokens_in":12238,"tokens_out":2919,"duration_ms":25540,"significance":"If the results are reproducible, the paper makes a useful contribution: it demonstrates that a flow-matching formulation with EMD pre-alignment can achieve competitive or better point cloud upsampling quality than diffusion baselines at a fraction of the sampling cost. The ablation 'Ours w/o EMD' (CD 2.817 vs 1.049 on PUGAN) gives clear evidence that the pre-alignment is important for the method's performance, and the extension to real-world RGB-D and LiDAR data is a strength. However, the paper's central reproducibility is undermined by an inconsistency between the stated Euler integrator and the sampling pseudocode, and by unspecified training hyperparameters. The absence of error bars or statistical tests also leaves several reported margins (e.g., PU1K 4x HD: 0.556 vs 0.583) without uncertainty quantification.","major_comments":[{"comment":"The sampling update in Algorithm 2, x_{t+1} = (1 - δ_t) x_t + δ_t ν_θ(x_t, t), is not the Euler step defined in Eq. (4), x_{t+1} = x_t + (δ_{t+1} - δ_t) ν_θ(x_t, t). These are two different discrete processes: the former includes an extra -δ_t x_t term and is not an Euler integration of the learned ODE. The manuscript does not state which update was used to produce Tables 1-4 and the 5-step runtime in Table 3. If Algorithm 2 was used, the reported results are not grounded in the flow-matching objective of Eq. (6); if Eq. (4) was used, the pseudocode is wrong and the method is underspecified without code. This is a correctness-level gap that must be resolved before the central claims can be accepted.","section":"Section 3.2, Eq. (4) vs Algorithm 2"},{"comment":"The paper trains the velocity field on EMD-pre-aligned pairs (Eq. 8, Algorithm 1) but at inference applies no alignment (Algorithm 2, Section 3.2). This is a distribution shift: the network must generalize from aligned training pairs to unaligned, possibly noisy or nonuniform sparse inputs. The manuscript does not analyze or quantify this shift, and the toy example in Figure 3 does not address it. Since the central claim is that PUFM 'directly learns the optimal transport between sparse and dense point cloud distributions,' the authors should provide evidence—for example, by comparing aligned vs unaligned inputs at inference, or by studying how the learned trajectories degrade with increasing misalignment—that the trained model transfers to raw unaligned inputs.","section":"Section 3.2, Eq. (8) and Algorithm 2"},{"comment":"In Table 4, the P2F values at noise level η=0.02 (e.g., PUDM 1.150, Ours 1.055) are about five times smaller than at η=0.01 (e.g., PUDM 6.025, Ours 5.887), while CD and HD degrade as expected with more noise. This internal inconsistency is not explained and undermines the robustness claim. The authors should either correct the metric values, explain why P2F behaves this way, or remove the affected rows.","section":"Section 4.5, Table 4"},{"comment":"The midpoint interpolation in Eq. (5) depends on two hyperparameters: the repeat factor γ and the noise level η. Neither is specified in the main text, and the supplementary is referenced only for the model architecture and sample schemes, not for these values. Without these values, the training-time source distribution is not fully specified, and the robustness experiments in Section 4.5 appear to use different η values than the main training setup. Please report the default η and γ used for training and testing.","section":"Section 3.2, Eq. (5)"}],"minor_comments":[{"comment":"The paper does not report error bars, standard deviations, or significance tests. Several margins in Table 1 are small (e.g., PU1K 4x HD: Ours 0.556 vs Grad-PU 0.583), so it would strengthen the comparison to include variance over multiple runs or seeds.","section":"Section 4.2, Table 1"},{"comment":"The reference style uses 'and et al.' for many entries, which makes it hard to identify authors and papers. Please use a standard bibliography format with full author lists or at least first-author names.","section":"Section 2 (References)"},{"comment":"The paper states 'we use Poisson disk sampling to generate uniform patches as ground truth' but does not specify the patch size or the exact protocol for extracting sparse inputs beyond randomly sampling 256 points. Providing the patch radius and sampling parameters would improve reproducibility.","section":"Section 4.1, Dataset description"},{"comment":"Equation (6) writes L(θ) = min_θ E ||...||, but the min is part of the optimization problem, not the loss definition. The equation should be L(θ) = E ||...||, or the min should be placed outside the expectation in the training objective.","section":"Section 3.2, Eq. (6)"}],"recommendation":"major_revision","confidential_remarks":"The strongest issue is the sampling update inconsistency, which is a correctness-level problem for reproducibility. I recommend requesting a revision that clarifies the integration scheme, specifies the missing hyperparameters, and corrects the Table 4 inconsistency. The paper's central idea is promising, and the ablation supports the pre-alignment claim, but the manuscript in its current form is not fully reproducible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the EMD pre-alignment trick is a genuine idea and the ablation supports it, but the paper has a load-bearing internal inconsistency in the sampling algorithm that has to be fixed before the numbers can be trusted.\n\nThe actual contribution is applying flow matching directly to sparse-to-dense point cloud upsampling, with midpoint interpolation to match cardinalities and EMD pre-alignment (training only) to stabilize the learned path. That is a reasonable technique-level contribution. The ablation is the strongest part: removing pre-alignment changes CD from 1.049 to 2.817 on PUGAN, and the toy experiment in Figure 3 supports the claimed mechanism. The efficiency story is also plausible in direction: 5 sampling steps vs 30 for PUDM is a meaningful reduction if it holds.\n\nNow the soft spots. The serious one is the sampling update. Equation 4 defines the Euler step as x_{t+1} = x_t + (δ_{t+1}−δ_t) νθ(x_t, t), which is the standard integration of dx/dt = νθ. Algorithm 2 says x_{t+1} = (1 − δ_t) x_t + δ_t νθ(x_t, t). These are not equivalent; Algorithm 2 is not an Euler step for the learned ODE. The paper never defines δ_t in Algorithm 2 and never states which update produced Tables 1–4 or the runtime in Table 3. If Algorithm 2 was used, the model is not integrating the flow it was trained on; if Eq. 4 was used, the pseudocode is wrong and the method is under-specified without code. Either way, the central claim is not reproducible from the manuscript as written.\n\nOther issues are more standard: η in Eq. 5 is never specified, no error bars or significance tests are given, several Table 1 margins are small, and the ScanNet/KITTI evidence is qualitative. The citation pattern looks fine.\n\nWho this is for: researchers in point cloud upsampling or efficient 3D generative models. The idea deserves a serious referee, but I would not cite it in its current form. I would send it to review with a request for a corrected algorithm, the missing hyperparameters, and code.","headline":"A real technique-level idea with a strong pre-alignment ablation, but the sampling update contradicts the paper's own Euler equation and must be fixed before the results are credible.","tokens_in":12749,"tokens_out":5562,"would_cite":false,"duration_ms":48029,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"PUFM learns a direct flow from sparse to dense point clouds, reporting state-of-the-art upsampling with only five sampling steps.","keywords":["point cloud upsampling","flow matching","Earth Mover's Distance","midpoint interpolation","diffusion models","generative model","3D vision","LiDAR point clouds"],"falsifier":"Compare PUFM's output on the same input processed with and without an EMD re-ordering at inference: if the Chamfer distance changes by more than the training-time margin, the learned flow depends on the training-time matching and will not transfer to arbitrary unaligned patches. A sweep of $\\eta$ in Eq. 5 across, say, 0.005 to 0.05 would also show whether the robustness claims hold outside an unspecified noise regime.","tokens_in":11752,"feed_emoji":"","tokens_out":9192,"duration_ms":74264,"temperature":0.7,"pith_summary":"This paper proposes PUFM, a flow-matching model for point cloud upsampling that maps a sparse input directly to a high-fidelity dense point cloud instead of first diffusing it into Gaussian noise. Midpoint interpolation densifies the sparse input to match the cardinality of the target, and an Earth Mover's Distance pre-alignment reorders the dense ground truth during training so the learned velocity field follows a stable straight path. On the PUGAN and PU1K benchmarks, the authors report better Chamfer, Hausdorff, and point-to-surface errors than prior diffusion and learning-based upsamplers, and they need only five Euler sampling steps compared with thirty for the diffusion baseline PUDM. The same model transfers to ScanNet RGB-D and KITTI LiDAR inputs, which is the practical payoff: fast, geometry-aware upsampling for real sensors.","feed_headline":"Sparse point clouds turn dense in five flow-matching steps","feed_subtitle":"By learning the direct sparse-to-dense path, PUFM beats diffusion baselines with far fewer sampling steps.","key_machinery":"The central object is the learned velocity field $\\nu_\\theta(x_t,t)$, trained so that integrating it transports the sparse-point distribution to the dense-point distribution. Three pieces make that training tractable: midpoint interpolation (Eq. 5) densifies the sparse cloud to match cardinality, EMD pre-alignment (Eq. 8) permutes the dense ground truth so the residual $x_1-\\tilde{x}_0$ is a coherent velocity target, and the straight interpolant $x_t=(1-t)\\tilde{x}_0+t x_1$ turns Eq. 6 into a simple regression on the displacement. A cosine schedule over $t$ concentrates training near small $t$, where the motion direction is easiest to pin down.","core_discovery":"PUFM's central claim is that the optimal transport between the sparse-point distribution and the dense-point distribution can be learned directly, without a noise-to-data detour. The paper parameterizes a velocity field $\\nu_\\theta(x_t,t)$ and trains it to reproduce the residual $x_1-\\tilde{x}_0$ along the straight interpolant $x_t=(1-t)\\tilde{x}_0+t x_1$, where $\\tilde{x}_0$ is the midpoint-interpolated sparse cloud. Because point clouds are unordered, the naive residual is ill-defined; the authors fix this with EMD pre-alignment (Eq. 8), which permutes the dense cloud to minimize the bijective assignment cost to the sparse cloud during training, and they show this removes the early-stage collapse seen without it. With the aligned objective, PUFM reports state-of-the-art CD, HD, and P2F on PUGAN and PU1K at 4x and 16x upsampling, and consistent gains over Grad-PU and PUDM across 5x to 32x scales. The efficiency claim is the five-step sampler: a single Euler pass from the sparse prior, rather than a long reverse diffusion chain.","pith_inferences":["The pre-alignment is a training-only crutch; a direct test would measure how output quality degrades as input patch ordering or noise shifts away from the EMD assignments seen in training.","Because the velocity field is defined for all $t\\in[0,1]$, the same midpoint-interpolation plus flow-matching recipe could apply to point cloud completion, denoising, or any unordered-set transport problem, not just upsampling.","The noise level $\\eta$ in Eq. 5 is never fixed in the main text, so the effective source distribution is under-specified; an ablation across $\\eta$ values would clarify whether the reported robustness is tied to a particular noise schedule."],"forward_implications":["If PUFM is correct, diffusion is not needed for point cloud upsampling: direct sparse-to-dense flow matching reaches better CD/HD/P2F with five Euler steps, versus thirty for PUDM.","The EMD pre-alignment is not optional: removing it raises PUGAN CD from 1.049 to 2.817 and PU1K CD from 0.545 to 1.188, confirming the matching stabilizes early training.","Because quality stays ahead across 5x to 32x factors, the method supports arbitrary upsampling by iterating 4x steps with FPS downsampling.","The model's robustness to Gaussian noise at $\\eta=0.01$ and $0.02$ and its qualitative results on ScanNet and KITTI imply the learned flow carries over to real RGB-D and LiDAR inputs."],"supporting_citations":[{"why":"Supplies the flow-matching objective, the straight-path interpolant, and the Euler integration recipe the method builds on.","marker":"[Lipman et al., 2023]"},{"why":"Provides the midpoint interpolation used to densify sparse inputs before flow matching.","marker":"[He and et al., 2023]"},{"why":"Supplies the EMD optimization procedure used to pre-align dense ground truth to sparse input.","marker":"[Chen and et al., 2020]"},{"why":"Provides the PointNet++ set-abstraction backbone that parameterizes the velocity network.","marker":"[Qi and et al., 2017]"},{"why":"The diffusion-based upsampling baseline (PUDM) that PUFM must beat in quality and sampling steps.","marker":"[Qu and et al., 2024]"},{"why":"Supplies the PUGAN benchmark and the adversarial upsampling baseline compared in Tables 1 and 2.","marker":"[Li et al., 2019]"},{"why":"Supplies the PU1K benchmark and the graph-convolution PUGCN baseline.","marker":"[Qian and et al., 2021]"},{"why":"Supplies furthest point sampling and the patch-based upsampling pipeline used for data preparation.","marker":"[Yifan and et al., 2019]"},{"why":"Provides ScanNet as the RGB-D real-world generalization test.","marker":"[Dai et al., 2017]"},{"why":"Provides KITTI as the LiDAR real-world generalization test.","marker":"[Geiger and et al., 2013]"}],"fun_headline_variants":["Flow matching turns sparse point clouds dense in five steps","No noise detour: PUFM upsamples point clouds with flow matching","EMD-stabilized flow matching beats diffusion for point cloud upsampling","Five-step Euler path wins over diffusion for point cloud densification"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the velocity field trained on EMD-pre-aligned synthetic pairs transfers to unaligned sparse inputs and to the noisy, nonuniform distributions found in real LiDAR and RGB-D data, even though pre-alignment is applied only during training and the noise level $\\eta$ in Eq. 5 is never specified.","fun_headline_variants_meta":{"raw":{"variants":["Flow matching turns sparse point clouds dense in five steps","No noise detour: PUFM upsamples point clouds with flow matching","EMD-stabilized flow matching beats diffusion for point cloud upsampling","Five-step Euler path wins over diffusion for point cloud densification"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000673,"raw_usage":{"total_tokens":3078,"prompt_tokens":976,"completion_tokens":2102,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":592,"completion_tokens_details":{"reasoning_tokens":2027}},"tokens_in":592,"tokens_out":2102,"duration_ms":13519,"temperature":1.0,"reasoning_tokens":2027,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:25:48.612450+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare PUFM's output on the same input processed with and without an EMD re-ordering at inference: if the Chamfer distance changes by more than the training-time margin, the learned flow depends on the training-time matching and will not transfer to arbitrary unaligned patches. A sweep of $\\eta$ in Eq. 5 across, say, 0.005 to 0.05 would also show whether the robustness claims hold outside an unspecified noise regime.","supporting_citations":[{"cited_title":"Pointmixup: Augmentation for point clouds","cited_arxiv_id":null,"evidence_quote":"Supplies the EMD optimization procedure used to pre-align dense ground truth to sparse input."},{"cited_title":"Pu-gcn: Point cloud upsampling using graph convolutional net- works","cited_arxiv_id":null,"evidence_quote":"Supplies the PU1K benchmark and the graph-convolution PUGCN baseline."},{"cited_title":"Yifan and et al","cited_arxiv_id":null,"evidence_quote":"Supplies furthest point sampling and the patch-based upsampling pipeline used for data preparation."},{"cited_title":"Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner","cited_arxiv_id":null,"evidence_quote":"Provides ScanNet as the RGB-D real-world generalization test."},{"cited_title":"Vision meets robotics: The kitti dataset","cited_arxiv_id":null,"evidence_quote":"Provides KITTI as the LiDAR real-world generalization test."}],"review_version":1}