{"id":"3f1456f3-679a-4923-b16d-903828629fed","arxiv_id":"2411.19036","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A pipeline that uses multi-view diffusion-generated depth images as shape priors, fused with the partial point cloud via attention and confidence filtering, achieves state-of-the-art completion on custom single-view benchmarks.","lead":"PCDreamer completes partial 3D scans by asking large multi-view diffusion models to generate images from other angles, then fusing those imagined views with the original scan. It reports better completion accuracy on synthetic benchmarks than prior geometry-only and image-guided methods, especially for thin or symmetric parts.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing control: the paper never replaces the diffusion-generated multi-view depth images with ground-truth renders or trivial views, so the central claim that diffusion priors supply the useful shape cues is not directly tested.","rationale":"The reader's weakest assumption is that the pipeline depends on generated multi-view depth images being view-consistent and shape-accurate; the paper's own earphone failure case shows the risk. My concern is adjacent but distinct: even if the generated images are reasonably accurate, the paper has not shown that the diffusion prior is the active ingredient in the measured improvement. The ablations vary the input modality but never vary the source of the multi-view images, so the improvement over partial-only input could come from added network capacity, from generic multi-view coverage, or from the specific semantic content of diffusion-generated views. Only the last interpretation supports the central claim. The missing control is inexpensive and would settle the causal question. The ShapeNet-34 unseen-category results and qualitative real-scan results provide some generalization evidence, but they do not address the causal role of the diffusion prior because the same generated-image pipeline is used throughout. The hyperparameter selection on the test set (Tabs. 7-8) and the custom synthetic evaluation protocol already justify the CONDITIONAL verdict; this missing control is an additional reason to maintain that verdict rather than upgrade to full acceptance. I do not see grounds for rejection: the method may well work as described, and the empirical improvements are internally consistent, but the central claim is underdetermined without the proposed control.","tokens_in":15445,"tokens_out":9484,"duration_ms":98222,"concrete_test":"Retrain the exact PCDreamer architecture with the same splits, training schedule, and losses, replacing the generated multi-view depth images in the fusion module with (a) ground-truth depth renders from the identical six camera poses, and (b) a capacity-matched non-semantic control (e.g., six copies of the input depth image, or the same images with random per-view geometric perturbations). Compare CD/DCD/F1 on the PCN test set. If (b) matches the reported 6.49, the improvement is not from diffusion priors; if (a) substantially beats the generated-images result, diffusion inconsistency is the bottleneck; only if generated images beat or match (a) and clearly beat (b) is the central claim supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is causal: multi-view diffusion priors provide global/local shape cues that improve completion. The ablations in Sec. 4.3 (Tabs. 4-5) only compare the full pipeline against variants that omit one modality or use RGB instead of depth; every multi-view condition is produced by the same ControlNet + Wonder3D/SVD + DepthAnything pipeline. There is no control in which the multi-view depth images are rendered from the ground-truth mesh at the same six poses, and no control with non-informative or perturbed images matched for model capacity. The reported gain (partial-only CD 6.91 vs. both 6.49, Tab. 4) could therefore be caused by the extra image encoder parameters, by simply having six views of the visible surface, or by the specific hallucinated content of the diffusion prior; only the last of these supports the paper's headline claim. The paper's own supplementary concedes that generated depth images exhibit inconsistencies, missing details, scale discrepancies, and noise, so the unconditional assertion that diffusion priors 'encode both global and local shape cues' is not established. This is a missing control, not a disagreement with the baseline numbers; it targets the interpretation of those numbers.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"PCDreamer proposes a three-stage pipeline for single-view point cloud completion: it first generates multi-view RGB and depth images of the imagined shape using frozen large diffusion models (ControlNet for depth-to-RGB, Wonder3D or Stable Video Diffusion for multi-view generation, and DepthAnything for depth estimation), then fuses these multi-view depth images with the partial point cloud in a patch-based attention fuser to produce a coarse point cloud, and finally applies a confidence-based point filtering and upsampling stage to obtain a dense, uniform completion. Only the fusion and consolidation modules are trained, optimized end-to-end with a hyperbolic Chamfer distance loss. The authors construct single-view variants of the PCN and ShapeNet-55 datasets, retrain ten prior methods from scratch on these custom benchmarks, and report improved average CD and DCD along with competitive F1 scores. Ablations cover the choice of visual modality, the fusion design, the confidence filtering, the number of views, and the filtering percentage, while the supplementary provides qualitative results on real scans and comparisons with SDS-Complete and single-image 3D generation methods.","tokens_in":15672,"tokens_out":11210,"duration_ms":96849,"significance":"If the central attribution holds, the contribution is a practical recipe for injecting geometric cues from frozen multi-view diffusion priors into a small, trainable completion network, together with a confidence-guided consolidation mechanism that tolerates the known inconsistencies of those priors. The paper has notable strengths: it retrains all baselines from scratch on the same data, evaluates two independent diffusion backbones (Wonder3D and SVD), ablates the main design choices, and includes an honest failure-case analysis. The consolidation idea could be of independent value for any method that consumes imperfect multi-view priors. However, the significance is currently bounded by three issues: the headline claim that the diffusion prior content itself provides the useful cues is not directly tested, one headline table contains a metric on which a baseline clearly wins, and all quantitative claims are made on self-constructed benchmarks rather than the standard test protocols. The stress-test concern about the missing control lands: it targets the interpretation of the numbers rather than the numbers themselves, and it can be answered with additional experiments.","major_comments":[{"comment":"The central claim that multi-view diffusion priors ‘encode both global and local shape cues’ that benefit completion is not directly tested by the current ablations. In every image-conditioned row of Tables 4 and 5, the multi-view depth images come from the same frozen chain (ControlNet + Wonder3D/SVD + DepthAnything); there is no condition in which these images are replaced by ground-truth renders at the same six camera poses, and no condition with non-informative images matched for encoder capacity. Since the partial-only row (CD 6.91, Table 4) removes the entire image encoder, the improvement to CD 6.49 with both modalities is also consistent with explanations that do not involve the diffusion prior’s content, such as added parameters or the mere availability of six additional views of the visible surface. This ambiguity is compounded by the depth-only ablation (CD 17.69), which shows that the raw generated depth images are nearly unusable on their own, and by the supplementary’s own list of defects in these images (inconsistencies, missing details, scale discrepancies, noise). I recommend adding two controls in Sec. 4.3: (i) the same fusion pipeline trained with ground-truth depth renders from the mesh at the six poses, and (ii) a non-informative image control (e.g., scrambled or blank depth images) to bound the contribution of the added encoder capacity. Without one of these controls, the paper’s headline attribution remains an interpretation rather than an established result.","section":"§3.1, §4.3, Tabs. 4–5"},{"comment":"Table 2 is in tension with the statement that the method ‘consistently outperforms other state-of-the-art approaches across all categories’ on ShapeNet-55. SVDFormer’s F-Score@1% is 0.447, which is clearly above Ours(SVD-MV) at 0.408 and Ours(Wonder3D) at 0.418, and the same pattern appears on the unseen categories of ShapeNet-34 in Table 3 (SVDFormer 0.348 vs. Ours(SVD-MV) 0.339). Since F1@1% is one of the three headline metrics the paper adopts, the text should acknowledge this trade-off and either explain it (for example, whether the confidence filtering removes points that contribute to the 1%-radius recall) or qualify the claim of uniform superiority. As written, the assertion of consistent superiority is contradicted by the paper’s own reported numbers.","section":"§4.2, Tab. 2"},{"comment":"All quantitative claims are made on a self-constructed single-view variant of PCN and ShapeNet-55, with a fixed viewpoint along the positive z-axis and back-projected partial point clouds. This is a legitimate design choice for studying the harder single-view setting, and retraining the baselines on the same data is fair internal comparison. However, because the test protocol differs from the standard PCN and ShapeNet-55 benchmarks, the reported numbers are not directly comparable with any published scores, and the claim of state-of-the-art performance rests solely on internally retrained baselines on the authors’ own split. I ask the authors to report results on the original benchmark protocols where feasible, to clearly frame the contribution as a new benchmark plus method otherwise, and to release the constructed datasets so that the comparison can be audited by the community.","section":"§4.1–§4.2"}],"minor_comments":[{"comment":"There are numerous typos and inconsistent notations that should be corrected: ‘Mutli-modality Fuser’ (Sec. 3.2), ‘baesd’ (Fig. 6 caption), ‘Poffet’ (Eq. 6), ‘Ff usion’ (Eq. 4), ‘0.0.617’ (Table 2, CRA-PCN DCD column), ‘the the image feature’ (Sec. 3.2), and the ungrammatical opening ‘PCDreamer, a novel algorithm’ in the introduction. Also, ‘we have proved in the ablation study’ (Sec. 3.2) should read ‘we show in the ablation study’.","section":"Throughout"},{"comment":"The definition of the confidence score is unclear: it is not specified over which dimension the average in Avg(dot(Q_i, K_i)) is taken, nor why a per-point self-attention dot product measures cross-view agreement. Please spell out the exact computation and the role of the sigmoid in producing the filtering percentage used in Sec. 4.4.","section":"§3.3, Eq. (5)"},{"comment":"The paper states that generated depth image values are normalized to the range [0, 1], but it does not say whether the normalization is per-view or joint across views. Since the supplementary acknowledges scale discrepancies between generated views, per-view normalization could distort relative depth cues; this choice should be specified and, if per-view, justified with an experiment or a comment.","section":"§3.1"},{"comment":"The generalization experiments on ScanObjectNN, KITTI, and randomly cropped point clouds are qualitative only. Given the introduction’s motivation of autonomous driving and robotics, a quantitative table on at least one real-scan set would substantially strengthen the claim of practical applicability.","section":"Supplementary, Fig. A1"}],"recommendation":"major_revision","confidential_remarks":"The skeptical reviewer’s missing-control concern is valid in my reading and should be answered with an experiment (ground-truth-render control and a non-informative-image control) rather than a rebuttal. The most actionable textual issue is the F1 contradiction in Table 2, which can be fixed by honest reporting and a short analysis. The custom benchmark is acceptable if released; I would endorse acceptance only after the attribution question is addressed and the dataset is made public for auditability. The paper fits the journal’s scope and the core direction is promising."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things. First, the paper is a competent integration of multi-view diffusion models into point cloud completion. Given a single-view partial scan, they render a depth image, use ControlNet to synthesize an RGB image, run Wonder3D or Stable Video Diffusion to generate six posed views, estimate depth with DepthAnything, then fuse image features with point cloud features via cross-attention and add a confidence-based consolidation step. That pipeline is new compared with prior image-guided completion methods, which need paired images or render depth from the partial cloud itself. The paper is clearly written and the ablations are honest: each module contributes, and the authors test two diffusion backbones and report a failure case in the earphone example.\n\nSecond, the central claim is not as well supported as it looks. The paper says the diffusion priors \"encode both global and local shape cues\" that drive the improvement. But there is no control in which the multi-view depth images are rendered from the ground-truth mesh at the same six poses, or in which the generated images are replaced with non-informative or perturbed views. So the reported gain of the full pipeline over partial-only (CD 6.49 vs 6.91) could come from the extra image encoder parameters, from having six views of the visible surface, or from the specific hallucinated content of the diffusion prior. Only the last of those supports the headline claim. The supplementary's own admission that generated depth images have inconsistencies, missing details, and scale discrepancies makes this under-determination concrete. This is a missing control, not a disagreement with the baseline numbers; it targets the interpretation.\n\nThat said, the comparison with SVDFormer, which uses multi-view depth from the partial cloud, partially mitigates the concern: PCDreamer outperforms it, so something is being added beyond simply seeing the visible surface from multiple angles. But a clean ground-truth-render control would have made the paper much stronger.\n\nWeaker points: the evaluation is on self-constructed variants of PCN and ShapeNet-55, not the standard partial-view protocols, and all baselines are retrained on the same custom data. That makes it hard to compare with any published numbers. No code or data are released. Hyperparameters like the number of views and filtering percentage appear to be chosen from the same test set, though the authors do not say this explicitly.\n\nWho this is for: people working on single-view point cloud completion and on using 2D diffusion priors for 3D tasks. It deserves a serious referee, but the referee should ask for the ground-truth render control, standard benchmarks, and at least a promise of code/data. Recommendation: engage with it, but treat the headline quantitative improvement as provisional.","headline":"Solid integration of diffusion priors for point cloud completion, but the key causal claim is under-tested by a missing control and the evaluation is on self-built datasets.","tokens_in":16204,"tokens_out":2404,"would_cite":true,"duration_ms":21797,"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":"This paper claims that multi-view images dreamed up by a diffusion model supply the missing geometric cues that let a single-view partial point cloud be completed with fine detail, and that fusing those images with attention plus…","keywords":["point cloud completion","multi-view diffusion priors","single-view partial point cloud","multi-modality fusion","depth image guidance","confidence-based consolidation","Chamfer distance","ShapeNet-55"],"falsifier":"Run PCDreamer on a set of partial point clouds built from slender, high-aspect-ratio objects such as earphones and thin tubular chair backs, and compare its Chamfer distance against the same model with the multi-view image branch removed; if the image-guided version is not systematically better on those cases, the claim that multi-view diffusion priors encode useful local cues fails exactly in the fine-detail regime the paper highlights.","tokens_in":15243,"feed_emoji":"🖼️","tokens_out":5658,"duration_ms":55484,"temperature":0.7,"pith_summary":"This paper argues that a partially scanned point cloud can be completed far more reliably when a large multi-view diffusion model is first used to 'dream' images of the object from viewpoints that were never observed. The central claim is that those synthetic multi-view images encode both global structure, such as symmetric parts hidden by self-occlusion, and local fine details like thin supports, and that this extra signal produces better completion than using the partial cloud alone. To test this, the authors build PCDreamer, a three-stage pipeline that generates posed multi-view depth images from a single-view depth image, fuses image and point-cloud features with attention, and then filters out unreliable points using a learned confidence score. On self-constructed single-view variants of the PCN and ShapeNet-55/34 benchmarks, PCDreamer reports the lowest Chamfer distance and highest F1-Score among the compared methods, with the largest gains on severe single-view occlusions.","feed_headline":"Diffusion-dreamed views fill in missing point cloud geometry","feed_subtitle":"Single-view scans lose half the shape; synthetic multi-view depth images recover symmetric parts and fine details.","key_machinery":"The load-bearing mechanism is the combination of a multi-view image generation pipeline and a two-stage shape refinery. The generation pipeline chains a depth-to-RGB translation model, a multi-view diffusion model, and a monocular depth estimator to turn one partial depth image into a set of posed multi-view depth images; this set is the imagined full-shape cue. The refinery consists of a multi-modality fuser, where a patch-based point-cloud encoder supplies queries and a pose-conditioned image encoder supplies keys/values in a cross-attention layer, followed by a shape consolidator that computes a per-point confidence score from self-attention over concatenated point-image features, filters the lowest 25 percent of points, and uses a per-point offset MLP to densify the remaining cloud. The confidence score is what lets the method tolerate the inevitable inconsistencies of diffusion-generated views.","core_discovery":"The paper's discovery is that multi-view diffusion priors, originally built for image-to-3D generation, can be repurposed as a free source of complete-shape cues for point cloud completion. Given a single-view partial point cloud, the method renders a depth image, converts it to RGB, asks a large multi-view diffusion model to generate several posed views of the full object, and estimates depth for each generated view. A patch-based Transformer encoder turns the partial point cloud into a query feature and the set of posed depth images into key/value features; cross-attention then produces a coarse complete point cloud. A confidence-guided consolidator scores each coarse point by its agreement with the multi-view image features and with other points, discards the low-confidence quarter, and upsamples the rest to a dense uniform completion. The authors report that this pipeline outperforms point-only and point-plus-image baselines on their datasets, especially for symmetric parts and fine structures that are invisible in the input view.","pith_inferences":["An upper-bound experiment the paper does not run: feeding ground-truth rendered multi-view depth images through the same fusion and consolidation modules would quantify exactly how much error is inherited from diffusion inconsistency versus from the fusion network itself.","A testable extension: make the filtering threshold per-point and view-aware instead of a fixed 75/25 split, which could rescue slender structures like the earphone failure the paper shows in Fig. 9.","The same fusion-plus-consolidation skeleton could be applied to LiDAR object completion in autonomous driving, using the viewpoint-selection strategy on real scans that the supplementary material begins to sketch."],"forward_implications":["If the central claim holds, single-view completion of severely occluded objects improves materially because the diffusion prior supplies symmetric counterparts and thin structures that geometric-only methods guess randomly.","The method is model-agnostic: swapping the multi-view generator (image-based versus video-based) changes results modestly, so better multi-view diffusion models should translate directly into better completion.","Confidence filtering acts as a safety valve: generated views that are locally inconsistent are down-weighted before densification, which is why the pipeline can tolerate imperfect dreams.","Because training uses only Chamfer distance on synthesized single-view partial clouds, the same recipe can be retrained on other completion benchmarks with paired partial and complete clouds."],"supporting_citations":[{"why":"Supplies the multi-view diffusion model that generates posed views of the imagined shape from a single RGB image.","marker":"[26]"},{"why":"Provides an alternative video-based multi-view diffusion backbone, used to show the pipeline is not tied to one generator.","marker":"[2]"},{"why":"Converts the single-view depth image into the RGB image that the multi-view diffusion generators require as input.","marker":"[52]"},{"why":"Estimates a depth image for each generated view, producing the 2.5D shape cue that the fuser consumes.","marker":"[48]"},{"why":"Supplies the PCN dataset and the canonical point cloud completion benchmark whose single-view variant the method is evaluated on.","marker":"[51]"},{"why":"Provides the hyperbolic Chamfer distance loss used to train the fusion and consolidation modules against noise and outliers.","marker":"[22]"},{"why":"Supplies the ShapeNet-55 benchmark and the PoinTr point-only baseline that the method is compared against.","marker":"[50]"},{"why":"Serves as one of the strongest recent point-only completion baselines that PCDreamer outperforms in average scores.","marker":"[31]"}],"fun_headline_variants":["Multi-view diffusion dreams complete point clouds","Dreaming up missing views to finish point clouds","Diffusion priors generate views to complete point clouds","Filling point cloud gaps with diffusion-view priors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline only completes as well as the imagined multi-view images are trustworthy: if the diffusion model generates inconsistent or geometrically wrong views (as it can for slender structures), the fusion and consolidation stages have no way to recover the true shape.","fun_headline_variants_meta":{"raw":{"variants":["Multi-view diffusion dreams complete point clouds","Dreaming up missing views to finish point clouds","Diffusion priors generate views to complete point clouds","Filling point cloud gaps with diffusion-view priors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000183,"raw_usage":{"total_tokens":1294,"prompt_tokens":907,"completion_tokens":387,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":523,"completion_tokens_details":{"reasoning_tokens":328}},"tokens_in":523,"tokens_out":387,"duration_ms":3897,"temperature":1.0,"reasoning_tokens":328,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:35:44.984740+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run PCDreamer on a set of partial point clouds built from slender, high-aspect-ratio objects such as earphones and thin tubular chair backs, and compare its Chamfer distance against the same model with the multi-view image branch removed; if the image-guided version is not systematically better on those cases, the claim that multi-view diffusion priors encode useful local cues fails exactly in the fine-detail regime the paper highlights.","supporting_citations":[{"cited_title":"Pcn: Point completion network","cited_arxiv_id":null,"evidence_quote":"Supplies the PCN dataset and the canonical point cloud completion benchmark whose single-view variant the method is evaluated on."},{"cited_title":"Hyperbolic cham- fer distance for point cloud completion","cited_arxiv_id":null,"evidence_quote":"Provides the hyperbolic Chamfer distance loss used to train the fusion and consolidation modules against noise and outliers."},{"cited_title":"Pointr: Diverse point cloud comple- tion with geometry-aware transformers","cited_arxiv_id":null,"evidence_quote":"Supplies the ShapeNet-55 benchmark and the PoinTr point-only baseline that the method is compared against."},{"cited_title":"Cra-pcn: Point cloud completion with intra-and inter-level cross-resolution transformers","cited_arxiv_id":null,"evidence_quote":"Serves as one of the strongest recent point-only completion baselines that PCDreamer outperforms in average scores."}],"review_version":1}