{"id":"dcb1f183-2f5d-4368-a1a3-e8d61dd36bba","arxiv_id":"2505.04622","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"PrimitiveAnything generates variable-length, multi-type primitive abstractions of 3D shapes with an auto-regressive transformer trained on 120K human-annotated assemblies, outperforming prior abstraction methods in geometric and perceptual metrics.","lead":"An auto-regressive transformer learns to generate 3D shapes as ordered lists of primitives (cuboids, elliptical cylinders, ellipsoids), trained on a new 120,000-shape dataset of human-crafted primitive assemblies. The authors report better geometric fidelity and human-perception alignment than optimization-based and earlier learning-based abstraction methods, with potential for lightweight, editable 3D content in games.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The perceptual claim rests on an unvalidated annotation ground truth: HumanPrim has no inter-annotator agreement analysis, and Appendix C concedes annotation style diversity; the model may be learning one annotation pool's conventions, not human decomposition generally.","rationale":"The stress-test pass searched for the weakest link in the chain from data to the perceptual claim. The strongest candidate is the unvalidated annotation ground truth: the paper's Appendix C admits annotation style diversity and over-segmentation, yet no inter-annotator agreement analysis is provided, and the segmentation metrics used as evidence of human alignment are computed against these same annotations. If the annotation signal is noisy or style-dependent, the model may simply reproduce one annotator pool's conventions, and the small user study cannot rescue that. A targeted re-annotation study would settle this directly. This agrees with the reader's weakest_assumption, so the verdict remains CONDITIONAL; the engineering contribution and reported ablations are real, but the perceptual claim needs the missing annotation-consistency evidence before it can be fully accepted.","tokens_in":18432,"tokens_out":7819,"duration_ms":86352,"concrete_test":"Recruit at least five independent annotators to re-decompose a stratified sample of 50 meshes drawn from HumanPrim (plus 20 held-out category meshes) using the same annotation tool and guidelines. Compute pairwise inter-annotator RI/VOI/SC and primitive-count distributions, then compute the same metrics between PrimitiveAnything outputs and each annotator. If the model-vs-annotator scores fall inside the annotator-vs-annotator envelope and that envelope is tight (e.g., pairwise RI above 0.9), the ground-truth assumption is supported; if annotator agreement is low, or if the model is no closer to individual annotators than those annotators are to each other, the 'human-perception alignment' claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that PrimitiveAnything 'better align[s] with human perception' depends on HumanPrim being a consistent, learnable signal of human decomposition. Appendix C explicitly concedes that 'some annotators tend to use more primitives than necessary ... resulting in over-segmentation in certain cases,' and that the annotation guidelines do not enforce a single abstraction level. No inter-annotator agreement metric is reported anywhere in the paper. The main quantitative evidence for human alignment, the RI/VOI/SC segmentation metrics in Tables 2 and 4, is computed by transferring ground-truth primitive labels to predicted primitives via nearest-neighbor point matching, and it is measured against this same unvalidated annotation pool. If annotator disagreement or style drift is substantial, the model is trained on a mixture of decomposition conventions, and the segmentation metrics only show consistency with one annotation pool's habits, not with human perception generally. The 30-participant, 20-shape user study in Appendix A is too small and too narrowly compared (only EMS and Marching-Primitives) to independently establish the perceptual claim. The paper's own limitation statement therefore flags the exact assumption that the central claim needs, and the paper provides no measurement showing that the human-decomposition signal is stable across annotators.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PrimitiveAnything, an autoregressive transformer that reformulates 3D shape primitive abstraction as a sequence-generation task. It introduces HumanPrim, a dataset of 120K human-crafted primitive assemblies; an ambiguity-free parameterization that unifies cuboids, elliptical cylinders, and ellipsoids; and a cascaded decoder that predicts primitive class, translation, rotation, and scale autoregressively. The model is conditioned on point clouds and trained with cross-entropy, EOS, and auxiliary Chamfer losses. Experiments compare against optimization-based methods (EMS, Marching-Primitives) and learning-based methods (Tulsiani et al. 2017, Paschalidou et al. 2019) on geometric and segmentation metrics, with ablations, a small user study, and qualitative generalization results. The paper claims that the method generates high-quality primitive assemblies that better align with human perception while maintaining geometric fidelity across diverse shape categories.","tokens_in":18673,"tokens_out":8600,"duration_ms":78224,"significance":"If the results hold, the work is significant: it provides a large-scale human-annotated dataset for primitive abstraction, a unified tokenization for multiple primitive types, and a generative formulation that can produce variable-length assemblies across categories. The reported geometric improvements over the considered baselines and the component ablations support the usefulness of the framework for reconstruction and editing. The main weakness is that the perceptual-alignment claim is not matched by the evidence: the annotation ground truth is not validated for inter-annotator consistency, and the primary quantitative evidence for 'human alignment' is measured against the same annotation pool used for training. The strengths are the dataset construction effort, the explicit ambiguity-free parameterization design with an ablation, and the inclusion of both geometric and segmentation metrics plus a user study.","major_comments":[{"comment":"The central claim of better alignment with human perception rests on the assumption that HumanPrim is a consistent, learnable signal of human decomposition, but the paper reports no inter-annotator agreement and Appendix C explicitly concedes that 'some annotators tend to use more primitives than necessary ... resulting in over-segmentation in certain cases.' Because the RI/VOI/SC metrics in Tables 2 and 4 are computed by transferring labels from the same HumanPrim annotations used for training, these numbers chiefly measure agreement with one annotation pool's conventions rather than with human perception generally. Please report pairwise annotator agreement on a double-annotated subset and validate the perceptual claim on independent annotations or a substantially larger user study.","section":"Section 4.1, Appendix C, Tables 2 and 4"},{"comment":"No quantitative table reports error bars, confidence intervals, or the number of independent runs or seeds over which the means are computed. Without this information, the reported margins (e.g., CD 0.0404 versus 0.0546 in Table 1) cannot be assessed for statistical significance. Please provide mean plus/minus standard deviation (or confidence intervals) and a repeated-seed or resampling protocol.","section":"Tables 1-5 and 7"},{"comment":"The user study uses 30 participants, 20 shapes, and only two optimization-based baselines, and it reports average scores without any dispersion or significance test. This is too narrow to independently support the perceptual-alignment claim made in the abstract. Please expand the study to more shapes and include at least one learning-based baseline, with per-participant variance and a significance test, or soften the perceptual claim accordingly.","section":"Appendix A, Table 6"},{"comment":"The generalization claim across diverse shape categories is supported quantitatively only on the ShapeNet chair category, which is a single-category cross-dataset test rather than a test of cross-category generalization, while the Objaverse evaluation is qualitative. Please provide quantitative held-out-category evaluation (e.g., a test split excluding categories seen in training) or an explicit analysis of train/test nearest-neighbor distances with numbers, to substantiate the 'across diverse shape categories' claim.","section":"Section 4.2 and Fig. 9"},{"comment":"The 'ambiguity-free' property is not proven. Equation (4) selects the symmetry rotation with minimal L1 norm, but the paper does not show that this arg min is unique across the discretized parameter space, and ties would reintroduce the mode confusion the scheme is designed to remove. Please prove uniqueness under the discretization or explicitly state and analyze the tie-breaking rule.","section":"Equations (3)-(6)"}],"minor_comments":[{"comment":"Equation (5) uses \\hat s_k, \\hat r_k, \\hat t_k on the left-hand side but s_k, r_k on the right-hand side; the composition order and notation should be clarified.","section":"Equation (5)"},{"comment":"Figure 3 contains confusing labels ('rb rbra', 'ra''); please redraw or clarify the caption and the text that refers to it.","section":"Figure 3"},{"comment":"The dataset name is misspelled as 'Objarverse' in Section 4.1 and the misspelling is repeated in Appendix A; please correct it to 'Objaverse'.","section":"Section 4.1"},{"comment":"The 'over 95% storage reduction' claim lacks a definition of the compared mesh representation and measurement protocol; please specify how the comparison was performed.","section":"Section 4.4"},{"comment":"The label-transfer procedure for RI/VOI/SC is described in one sentence; please provide explicit steps, including how empty or duplicate predicted primitives are handled.","section":"Section 4.1"},{"comment":"'We employs the pointbert-vitg14' contains a grammatical error; the retrieval experiment would also be stronger if reported with quantitative similarity scores rather than qualitative examples only.","section":"Appendix A, Generalization Analysis"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nFirst thing to know: this is a real step forward for primitive-based 3D abstraction. The paper reformulates shape abstraction as auto-regressive sequence generation over cuboids, elliptical cylinders, and ellipsoids, trained on 120K human-annotated decompositions (HumanPrim). The multi-type unified parameterization, with the symmetry-aware ambiguity-free scheme, is genuinely new and sensible. The geometric results are good: it beats Marching-Primitives and EMS on CD/EMD/Voxel-IoU, and it generalizes to ShapeNet chairs without training on that dataset.\n\nWhat does it do well? The ablations are clean, and each component — ambiguity-free parameterization, cascaded decoding, auxiliary Chamfer loss — shows a positive effect. The pipeline is well engineered, and Appendix C is candid about limitations: over-segmentation, discretization precision loss, missing symmetry integration, and annotation style variation. That honesty counts.\n\nWhere are the soft spots? The headline claim is “better align with human perception,” and that is the weakest part of the evidence. The segmentation metrics (RI/VOI/SC) are computed by matching predicted primitives to ground truth via nearest-neighbor label transfer, and that ground truth is the HumanPrim annotation pool itself. No inter-annotator agreement is reported. Appendix C explicitly admits some annotators over-segment and that abstraction level varies. So those metrics mostly show the model reproduces one annotation pool’s conventions, not stable human decomposition preferences. The user study is independent, but 30 participants and 20 shapes, with only EMS and Marching-Primitives as comparisons, is thin; no variance or significance numbers are given. There are also no error bars on any quantitative table, and the test set is limited to 314 “high quality” artist-labeled samples — selection bias that favors the method. Code and dataset are not released, which hurts reproducibility.\n\nNone of this sinks the paper. The engineering and geometric results stand on their own. But the perceptual claim needs a stronger foundation: inter-annotator agreement analysis, a larger independent user study, and error bars. Those are fixable in revision.\n\nWho is this for? Anyone working on shape abstraction, interpretable 3D representations, or auto-regressive generation for geometric structures. It deserves a serious referee — send it out. I would set expectation of major revision.","headline":"A solid auto-regressive framework and a large human-annotated primitive dataset, but the human-perception claim rests on thinner evidence than the paper suggests.","tokens_in":19206,"tokens_out":2338,"would_cite":true,"duration_ms":23283,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A shape-conditioned transformer trained on 120K human-annotated primitive assemblies can regenerate a broad range of 3D objects as sequences of cuboids, elliptical cylinders, and ellipsoids, with geometric fidelity close to…","keywords":["3D shape abstraction","primitive assembly generation","auto-regressive transformer","human-annotated decompositions","shape decomposition","point-cloud conditioning","primitive parameterization","3D content generation"],"falsifier":"Run a forced-choice user study on fresh shapes from categories absent from HumanPrim, where participants compare PrimitiveAnything's output against human-authored primitive assemblies, and also measure inter-annotator agreement on those same shapes; if participants do not significantly prefer the human assemblies over the model's, or if pairwise human agreement is low, then the central claim that the model reproduces human decomposition patterns is falsified.","tokens_in":18230,"feed_emoji":"🧱","tokens_out":6054,"duration_ms":56180,"temperature":0.7,"pith_summary":"The paper tries to establish that shape primitive abstraction—breaking a 3D mesh into simple geometric elements—can be recast as an autoregressive sequence-generation task learned from human-annotated examples. The payoff would be a single primitive-based representation that transfers across many object categories, is editable by non-experts, and is compact enough for game-ready user-generated content. The proposed model, PrimitiveAnything, directly learns the order and parameters of primitive assembly from 120K human-crafted decompositions, rather than optimizing geometric fit alone, and claims to produce assemblies that both match geometry and align with how people perceive and build shapes.","feed_headline":"Transformer turns 3D shapes into human-style primitive assemblies","feed_subtitle":"Learned from 120K hand-crafted decompositions, it beats optimization baselines on fidelity and human alignment.","key_machinery":"The central mechanism is the ambiguity-free parameterization scheme: each primitive is represented by a type label plus translation, rotation in Euler angles, and scale, and symmetry-equivalent parameter tuples are canonicalized by composing the primitive's rotational symmetry set with the current transform and selecting the tuple with minimal L1 rotation norm. This removes the one-to-many ambiguity that symmetry creates—for example, a cuboid's 90-degree rotation combined with an axis swap of scale factors produces an identical shape—so the autoregressive transformer is trained on a stable sequence-to-primitive mapping. The architecture couples a primitive encoder, a decoder-only transformer, a cascaded primitive decoder that predicts class, translation, rotation, and scale in that order, and an EOS decoder, with Michelangelo point-cloud tokens serving as the conditioning signal.","core_discovery":"On its own terms, PrimitiveAnything claims that a shape-conditioned decoder-only transformer, trained on the HumanPrim dataset of 120K manually assembled primitive decompositions, can generate variable-length sequences of primitives from three types—cuboids, elliptical cylinders, and ellipsoids—whose union reconstructs the input shape. On the HumanPrim test set, the method reports lower Chamfer distance and Earth Mover's distance than optimization-based baselines, higher Voxel-IoU, and better scores on the segmentation metrics RI, VOI, and SC, which the paper interprets as closer alignment with human decomposition patterns. The model also produces plausible primitive assemblies for shapes from Objaverse and for ShapeNet chair categories it was not trained on, suggesting that the learned assembly logic generalizes beyond the training distribution.","pith_inferences":["The same canonicalization recipe could be applied to richer primitive vocabularies—superquadrics, convex polytopes, or CSG operators—because the framework treats primitive type as a token and keeps the sequence mechanics unchanged.","The z-y-x centroid sorting of primitives is a design choice; testing alternative orderings (by size, by semantic part, or by connected component) could reveal how much of the learned 'human logic' actually resides in the ordering versus the transformer's predictive capacity.","A direct test of the human-alignment claim would be an inter-annotator agreement study on fresh meshes: without a high baseline agreement among human annotators, the ground-truth 'human perception' is itself an unstable target, regardless of model scores.","The paper's storage-efficiency and editability claims could be quantified further by measuring how much a user must manipulate the generated primitives to achieve a target edit, which the paper does not currently report."],"forward_implications":["Primitive abstraction becomes a scalable generation problem: a single model can output variable-length assemblies of multiple primitive types from an arbitrary point cloud, without per-category training.","The resulting representations are compact and editable—the paper reports over 95% storage reduction relative to meshes—which could enable primitive-based user-generated content in games and simple manipulation in standard graphics interfaces.","Because primitive type is handled as a learnable token, the framework can be extended to new primitive types without architectural changes, making the approach adaptable to richer primitive vocabularies.","When chained with text-to-3D or image-to-3D generators, the method can produce primitive-based 3D content directly from natural user conditions, as demonstrated with TRELLIS and SDXL in the paper.","The improved segmentation metrics indicate that the assemblies carry part-level structure useful for downstream recognition or editing tasks, not just surface fidelity.","The method's generalization across Objaverse suggests that human assembly logic, as captured by the annotations, transfers to shape categories beyond the training set."],"supporting_citations":[{"why":"MeshAnything's autoregressive artist-mesh generation paradigm motivates reformulating shape abstraction as sequence generation.","marker":"[Chen et al. 2024a,b]"},{"why":"The Michelangelo encoder converts input point clouds into fixed-length token sequences used as conditioning for the primitive transformer.","marker":"[Zhao et al. 2023]"},{"why":"Superquadric-based learning method serves as a representative baseline that PrimitiveAnything compares against and extends by using multiple primitive types.","marker":"[Paschalidou et al. 2019]"},{"why":"Volumetric-primitive assembly learning method is a baseline on chair categories, representing the cuboid-only approach that PrimitiveAnything outperforms.","marker":"[Tulsiani et al. 2017]"},{"why":"Marching-Primitives is a state-of-the-art optimization-based baseline using superquadrics, compared on geometry and segmentation metrics.","marker":"[Liu et al. 2023b]"},{"why":"EMS is a probabilistic superquadric fitting baseline that the paper uses as a geometric and perceptual comparison.","marker":"[Liu et al. 2022]"},{"why":"Defines the Chamfer distance used as an auxiliary geometric loss to guide per-primitive reconstruction fidelity.","marker":"[Fan et al. 2017]"},{"why":"Gumbel-Softmax enables differentiable sampling of discrete primitive attributes during training of the autoregressive model.","marker":"[Jang et al. 2017]"},{"why":"Objaverse supplies out-of-distribution shapes for qualitative generalization evaluation and the user study.","marker":"[Deitke et al. 2023]"}],"fun_headline_variants":["PrimitiveAnything: Transformer learns human-style 3D shape decomposition","Auto-regressive transformer assembles 3D shapes from primitives","Learning primitive assembly from 120K human-crafted shapes","Transformer beats optimization on 3D shape abstraction fidelity","Human-crafted 3D decompositions teach Transformer to generalize"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the HumanPrim annotations are a consistent, learnable ground truth for 'human perception' of shape decomposition; the paper itself acknowledges in the limitations that annotation styles vary and some assemblies over-segment, and no inter-annotator agreement statistic is reported, so if the annotations are idiosyncratic or do not transfer to unseen categories, the perceptual-alignment claim loses its foundation.","fun_headline_variants_meta":{"raw":{"variants":["PrimitiveAnything: Transformer learns human-style 3D shape decomposition","Auto-regressive transformer assembles 3D shapes from primitives","Learning primitive assembly from 120K human-crafted shapes","Transformer beats optimization on 3D shape abstraction fidelity","Human-crafted 3D decompositions teach Transformer to generalize"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000664,"raw_usage":{"total_tokens":3018,"prompt_tokens":916,"completion_tokens":2102,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":532,"completion_tokens_details":{"reasoning_tokens":2017}},"tokens_in":532,"tokens_out":2102,"duration_ms":13373,"temperature":1.0,"reasoning_tokens":2017,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:23:05.036175+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a forced-choice user study on fresh shapes from categories absent from HumanPrim, where participants compare PrimitiveAnything's output against human-authored primitive assemblies, and also measure inter-annotator agreement on those same shapes; if participants do not significantly prefer the human assemblies over the model's, or if pairwise human agreement is low, then the central claim that the model reproduces human decomposition patterns is falsified.","supporting_citations":[],"review_version":1}