{"id":"86c31bb2-a990-4774-8c09-7d013def59d4","arxiv_id":"2506.17074","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Assembler casts 3D part assembly as a diffusion process over sparse anchor point clouds, achieving state-of-the-art on PartNet and extending to general, image-conditioned objects.","lead":"A new computer vision system, Assembler, takes a set of 3D object parts plus a reference image and produces a complete assembled 3D model by generating a cloud of anchor points that show where each part belongs. It reports state-of-the-art results on PartNet and claims the first automatic part assembly for diverse everyday objects.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"In-the-wild claim rests on qualitative Toys4K examples; no quantitative metrics or comparison baseline are reported for the general-object setting.","rationale":"The reader's weakest_assumption correctly flags the synthetic data pipeline and the two-anchor underdetermination, but the most decision-relevant issue for the paper's headline claim is the lack of any quantitative evaluation in the general-object regime. The PartNet results are competently reported and the ablation in Table 2 is informative, but PartNet is category-specific and not the setting where the paper claims to be first. The Toys4K experiments are the only direct evidence for the in-the-wild claim, and they are purely qualitative. A concrete quantitative evaluation with a strong image-only baseline would settle whether the diffusion model is actually learning assembly structure or mostly copying the reference image. This does not change the reader's CONDITIONAL verdict: the concern is substantive but addressable, and the method has credible components. I therefore keep the verdict unchanged while sharpening the specific test needed to support the strongest claim.","tokens_in":15757,"tokens_out":5302,"duration_ms":58626,"concrete_test":"Evaluate the released Assembler model on the Toys4K test split against the original object meshes as ground truth, reporting SCD, PA, CA, and SR over at least three diffusion samples per object. Include two baselines: (1) an image-only registration method that aligns input parts to the reference image using DINOv2 patch features, and (2) a category-agnostic pose predictor such as SPAFormer retrained on the same synthetic dataset. If Assembler's PA/SR advantage over the image-only baseline is small or within noise, the in-the-wild generalization claim is not established; if the advantage is large, the claimed contribution is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.3 states that Assembler 'achieves, for the first time, reasonable 3D part assembly of general objects', and the conclusion claims 'high-quality, automatic 3D part assembly for diverse, in-the-wild objects'. The only evidence offered for this central claim is a set of qualitative images (Figures 4 and 7). No SCD, PA, CA, or SR numbers are reported for the Toys4K test set, and no non-trivial baseline is compared in this setting. Because Toys4K objects are segmented with the same connected-component splitting, KNN grouping, and random rigid perturbation pipeline used to create the training data, the experiment primarily measures the model's ability to invert known synthetic perturbations, not necessarily its ability to handle broken, incomplete, interpenetrating, or non-rigid real parts. Moreover, the reference image shows the original assembled object, so an image-only or retrieval-based method could plausibly account for much of the apparent success; without such a baseline, the 'high-quality, in-the-wild' claim is not evidenced. This is the load-bearing gap: the strongest claim of the paper is supported only by unquantified, subjective examples in exactly the regime where the method is claimed to be novel. A separate technical issue in Section 3.1, that two anchor points do not determine a rigid transformation, is real but mitigated by the stated minimum of ten points per part in Appendix B, so it is not the primary risk to the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Assembler, a diffusion-based framework for 3D part assembly from input part meshes and a reference image. Instead of predicting per-part SE(3) poses, it represents each part by a sparse anchor point cloud and generates the assembled object's anchor points in Euclidean space via a DiT model, then recovers rigid transforms by least-squares fitting. The authors additionally curate a 320K-object synthetic assembly dataset from several mesh repositories and report state-of-the-art results on the PartNet benchmark. They further demonstrate a prototype part-aware 3D generation pipeline combining a vision-language model, TripoSG, and Assembler. The central novelty claim is that Assembler is the first to achieve high-quality, automatic assembly of general in-the-wild objects, supported by qualitative results on Toys4K.","tokens_in":15993,"tokens_out":4731,"duration_ms":46006,"significance":"If the claims are supported, the anchor point representation is a sensible way to sidestep SE(3) generative modeling, and the large-scale synthetic data pipeline is a potentially valuable resource for the community. The PartNet results are strong on the connectivity and success-rate metrics, and the ablations support the importance of the concatenation-based conditioning and the anchor point budget. The PartNet evaluation is quantitative with four metrics and ablations, and the authors promise code release; however, the general-object claim is not yet backed by comparable quantitative evidence, which is the main gap.","major_comments":[{"comment":"The central claim of generalization to in-the-wild objects is supported only by qualitative examples; no SCD, PA, CA, or SR numbers are reported for the Toys4K test set, and no non-trivial baseline is compared in this setting. Because Toys4K objects are processed with the same connected-component splitting, KNN grouping, and random rigid perturbation pipeline used to create the training data (§3.3 and Appendix A), the experiment predominantly measures the model's ability to invert synthetic perturbations drawn from the training distribution. Please add quantitative evaluation on Toys4K with at least one baseline (e.g., a retrieval-based or registration-based method), and consider a held-out test set segmented by an independent protocol to break this circularity. This is load-bearing for the paper's strongest claim.","section":"§4.3, Figures 4 and 7"},{"comment":"The text states that 'each part at minimum requires two anchor points to compute the transformation.' Two point correspondences do not determine a rigid transformation in SE(3); they leave a one-parameter family of rotations about the line connecting the points, so the least-squares fit is underdetermined. Appendix B later states that a minimum of 10 anchor points per part is used, which is sufficient, but the main-text justification should be corrected to state the true minimal requirement (three non-collinear points) and explain why 10 is used in practice.","section":"§3.1"},{"comment":"The comparison is not fully controlled: the 'Ours-img' row uses a reference image as an additional condition, whereas DGL, RGL, Score-PA, and SPAFormer do not receive any image input. Because the image provides a strong disambiguation cue, the gains of Ours-img over these baselines could be due to the extra input rather than the proposed architecture. The 'Ours' row is the fair comparison and does show improvements in PA, CA, and SR, but its SCD is worse; the text should explicitly acknowledge this and state which comparison is used to support the state-of-the-art claim.","section":"§4.2, Table 1"},{"comment":"The part-aware 3D generation prototype is demonstrated on a single qualitative example without quantitative metrics or comparisons to existing part-aware generation methods. The claim of 'high-quality, editable objects' is therefore not yet supported. Either provide a small user study or quantitative evaluation, or explicitly frame this section as a proof-of-concept.","section":"§4.5, Figure 5"}],"minor_comments":[{"comment":"The notation 'S= —N i P′i' is malformed; it should express a union or concatenation of the transformed parts.","section":"§3.1"},{"comment":"There is a typo in the introduction: 'constrast' should be 'contrast'.","section":"§1"},{"comment":"There is a typo: 'probablistic' should be 'probabilistic'.","section":"§2.1"},{"comment":"Several important hyperparameters, such as the noise schedule, classifier-free guidance scale, and training resolution, are not given in the main text and are deferred to the supplementary material; consider stating them briefly in the main text for reproducibility.","section":"§4.1"},{"comment":"In the printed figures, the 'Input Parts' and 'Anchor Points' columns are difficult to compare because part colors are not consistently mapped across columns; add a color correspondence or explicit labels.","section":"Figures 4 and 7"},{"comment":"The concatenation notation 'xt#cp' is nonstandard and could be confused with convolution; use 'concat' or a clearer operator such as ⊕.","section":"Eq. (1)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript presents a promising method with a novel representation and a large-scale data pipeline, and the PartNet results are solid. However, the central claim of in-the-wild generalization rests entirely on qualitative evidence, and the evaluation design has a same-distribution circularity. I would advise the editor to require quantitative results on Toys4K and preferably an independent test set before considering acceptance. The image-conditioned comparison in Table 1 also needs clarification. The paper is otherwise well written and within the scope of a leading vision or graphics venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The anchor-point diffusion idea is genuinely good, and the PartNet results are solid. But the headline claim about in-the-wild assembly is supported only by qualitative examples, so don't take it at face value. The representation is the real contribution: it avoids SE(3) pose estimation, handles varying part counts, and makes generative assembly scalable. The PartNet experiments are solid—they beat four open baselines on PA, CA, and SR across Chair/Table/Lamp, and the ablations (cross-attention failure, CFG, anchor count, data diversity) support the design choices. The 320K-object synthesis pipeline is simple but clever, and even if the dataset isn't released yet, the idea is reproducible.\n\nThe soft spot is the central claim. Section 4.3 and the conclusion say Assembler is 'the first' to achieve 'high-quality, automatic 3D part assembly for diverse, in-the-wild objects.' The evidence for that is Figures 4 and 7 only—qualitative Toys4K examples, no metrics, no baseline. Toys4K is processed with the same connected-components/KNN/random-perturbation pipeline used for training, so the experiment mostly tests whether the model can invert the known augmentation. A trivial baseline—say, retrieving the nearest training shape or registering parts to the reference image—might do surprisingly well, and there's no way to tell from the paper. The two-anchor-points remark in Section 3.1 is mathematically wrong (two point correspondences leave a rotational DOF), but Appendix B says the actual minimum is 10 points per part, so it's a sloppy sentence rather than a real flaw. Minor, but worth fixing.\n\nThe absence of code and data at submission time also hurts reproducibility, though they say it will be released. The part-aware generation prototype (GPT-4o + TripoSG + Assembler) is a nice demo but only one example; I wouldn't weigh it heavily either way.\n\nOverall: the core idea is plausible, the PartNet results are credible, and the limitations section is honest. The paper deserves peer review, and I'd send it back with a request for quantitative general-object evaluation, at least one non-trivial baseline in that setting, and a discussion of where the pipeline could fail on real, broken, or interpenetrating parts. I'd cite the anchor-point representation and the PartNet results, but not the in-the-wild claim as it stands.","headline":"The anchor-point diffusion idea is genuinely good and the PartNet results are solid, but the in-the-wild claim rests on qualitative examples only. Worth a serious referee, with the general-object evaluation needing real work.","tokens_in":16543,"tokens_out":5061,"would_cite":true,"duration_ms":43328,"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":"Assembler claims that general 3D part assembly is best treated as diffusion over sparse anchor point clouds, and backs the claim with a 320K-object synthetic dataset and category-level gains on PartNet.","keywords":["3D part assembly","diffusion models","sparse anchor points","point cloud generation","part-aware 3D modeling","generative modeling","synthetic dataset","PartNet"],"falsifier":"Run the trained model on objects whose parts were obtained by physically separating or scanning real components, so the parts have interpenetration, gaps, or missing geometry; if part accuracy drops far below the synthetic test result, the connected-component synthesis assumption is falsified.","tokens_in":15521,"feed_emoji":"🧩","tokens_out":7843,"duration_ms":73795,"temperature":0.7,"pith_summary":"The paper tries to show that general 3D part assembly can be solved at scale by rephrasing it as generative point-cloud diffusion rather than per-part pose prediction. Given part meshes and a reference image, the model generates the assembled object's sparse anchor points directly in Euclidean space, then recovers each part's pose by least-squares fitting. A 320K-object synthetic dataset built by splitting artist meshes into connected components supplies the scale needed for category-agnostic training. If the approach holds, category-specific assembly networks can be replaced by a single model that handles varied part counts, repeated parts, and unseen object categories, and the paper reports state-of-the-art PartNet results plus qualitative in-the-wild assemblies.","feed_headline":"Diffusion over anchor points makes 3D assembly general","feed_subtitle":"A 320K-object synthetic dataset lets one model assemble chairs, tables, toys, and unseen in-the-wild objects.","key_machinery":"The load-bearing device is the sparse anchor point cloud with a fixed total budget of 1024 points distributed across parts by size ratio, paired with a diffusion transformer (DiT) denoiser that concatenates per-part conditions directly to the noised point tokens. Each part condition combines its original anchor coordinates, shape latents from the Dora VAE, and a part-index Fourier embedding, while a block-diagonal attention mask keeps tokens attending only within their own part. Reference-image information enters through DINOv2 features via cross-attention. The generated point cloud itself is the output representation, and poses are obtained by least-squares fitting, so the network never explicitly predicts rotations; the paper also abandons cross-attention for part conditions because it breaks the per-point alignment needed to preserve rigidity and point order.","core_discovery":"Assembler claims that part assembly is better posed as conditional generation of an assembled sparse point cloud than as deterministic SE(3) pose regression. Each input part is sampled into anchor points; the diffusion model outputs the anchors of all parts placed in object coordinates, and least-squares rigid fitting converts those anchors into per-part transformations. Because generation happens over point positions rather than rotations, the model can represent several valid configurations at once and does not need a fixed category or part count. Trained on over 320K synthesized assemblies, the model outperforms prior category-specific methods on PartNet, and the paper presents it as the first demonstration of high-quality automatic assembly for diverse real-world objects.","pith_inferences":["A natural extension is assembly from damaged or partial parts, since the training pipeline assumes clean connected components; simulating interpenetration, gaps, or missing geometry at training time could test whether the anchor-point representation survives real-world part conditions.","The anchor-point format could serve as a generic interchange representation for part-aware 3D generation, letting any image-to-3D or text-to-3D generator whose output is segmented be assembled without retraining the diffusion model.","The model's per-part generation confidence could double as a diagnostic for input segmentation quality: parts whose predicted anchors are inconsistent are likely mis-grouped or semantically split by the data pipeline.","If the Euclidean anchor-point formulation scales, fractured-object reassembly might also be reframed as anchor-point diffusion, though tight boundary alignment would likely demand far more than 1024 anchors per object."],"forward_implications":["A single Assembler model can replace category-specific assembly networks, since the same architecture handles chairs, tables, lamps, and objects with 3 to 100 parts.","The fixed anchor-point budget decouples model size from part count, so objects with many small parts remain tractable as long as every part receives enough points to define its transform.","Because the output is a sampled distribution, repeated and symmetric parts can be resolved probabilistically, and the reference image further disambiguates which configuration is intended.","Category-agnostic training helps individual categories: the paper's ablation shows that adding non-chair PartNet data improves chair assembly even though the extra data contains no chairs.","The same model becomes the bottom-up stage of a part-aware generation pipeline, turning image-derived part meshes into a complete, editable object."],"supporting_citations":[{"why":"Supplies the PartNet benchmark and part-level annotations used for category-specific evaluation and as an additional training source.","marker":"[Mo et al. 2019]"},{"why":"Provides Objaverse meshes whose connected components become a large share of the synthesized 320K assembly dataset.","marker":"[Deitke et al. 2023b]"},{"why":"Supplies the TRELLIS-500K mesh collection that the data pipeline filters, segments, groups, and augments.","marker":"[Xiang et al. 2024]"},{"why":"The Dora VAE encodes each part's dense point cloud into shape latents attached to the sparse anchor points.","marker":"[Chen et al. 2024b]"},{"why":"The DINOv2 image encoder extracts reference-image features injected by cross-attention into the diffusion transformer.","marker":"[Oquab et al. 2023]"},{"why":"Prior score-based generative part assembly that motivates the generative formulation Assembler extends.","marker":"[Cheng et al. 2023]"},{"why":"SPAFormer is the strongest prior baseline and supplies the processed PartNet data format used for fair comparison.","marker":"[Xu et al. 2025b]"},{"why":"DGL defines the deterministic graph-based pose-prediction paradigm that Assembler contrasts with its generative anchor-point approach.","marker":"[Zhan et al. 2020]"}],"fun_headline_variants":["Anchor point diffusion scales 3D part assembly","Diffusion over anchors assembles any object","From parts and a reference image, diffusion assembles","General 3D assembly with anchor point diffusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The approach assumes that real-world parts are well represented by clean, complete, semantically grouped pieces obtained by splitting artist meshes into connected components and perturbing them, so that the synthetic training distribution matches the intended test scenarios.","fun_headline_variants_meta":{"raw":{"variants":["Anchor point diffusion scales 3D part assembly","Diffusion over anchors assembles any object","From parts and a reference image, diffusion assembles","General 3D assembly with anchor point diffusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00086,"raw_usage":{"total_tokens":3722,"prompt_tokens":922,"completion_tokens":2800,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":538,"completion_tokens_details":{"reasoning_tokens":2741}},"tokens_in":538,"tokens_out":2800,"duration_ms":18453,"temperature":1.0,"reasoning_tokens":2741,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:11:59.418701+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained model on objects whose parts were obtained by physically separating or scanning real components, so the parts have interpenetration, gaps, or missing geometry; if part accuracy drops far below the synthetic test result, the connected-component synthesis assumption is falsified.","supporting_citations":[],"review_version":2}