{"id":"fd48b1d3-11ee-4779-9001-09622d4107dd","arxiv_id":"2608.13541","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"SCULPT generates a complete 3D object together with its parts by repeatedly predicting one extracted part and an updated remainder, achieving state-of-the-art part geometry on PartObjaverse.","lead":"SCULPT turns a single 2D image into a complete 3D object that is already divided into meaningful parts, using a subtractive loop that peels off one part at a time while keeping the rest. A generalist should read it because editable, part-level 3D generation is a bottleneck for game assets, AR, and product design.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"State-of-the-art claim is not supported by the evaluated baseline set: the two closest methods (UniPart, AutoPartGen) are absent, and margins over the strongest included baseline are small and unquantified.","rationale":"I read the paper as making a strong empirical claim: SCULPT produces the best part, semantic-group, and object geometry among part-aware generators on PartObjaverse. The method itself is clearly motivated, and the ablations in Table 2 provide useful support for the individual components. My concern is not with the internal mechanism but with whether the evidence as presented can carry the word 'state-of-the-art.' The paper explicitly distinguishes UniPart and AutoPartGen as the closest prior formulations, so omitting them from Table 1 is a material gap. The small differences from TRELLIS.2+PartField make the omission especially consequential: without paired uncertainty estimates, a 0.0001 CD gap cannot be distinguished from evaluation noise. I partially agree with the reader's verdict: the reader also noted the missing baselines and lack of statistical confidence in the rationale, although the stated weakest assumption focused on latent capacity for sub-objects. That latent-capacity concern is plausible but more speculative; the end-to-end metrics already exercise the full recurrent pipeline, whereas the missing-baseline issue directly tests whether the central superiority claim is true. A direct comparison with UniPart and AutoPartGen, plus paired confidence intervals, would settle the concern. This reinforces, rather than changes, the conditional verdict.","tokens_in":17981,"tokens_out":14788,"duration_ms":171790,"concrete_test":"Obtain official UniPart and AutoPartGen checkpoints and run them on the same 200 PartObjaverse images with the same conditioning inputs and the same Hungarian-matched CD/F1 code used for Table 1. Report paired per-object differences and 95% bootstrap intervals for SCULPT versus each baseline at the part, semantic-group, and object levels. If any baseline's interval overlaps or beats SCULPT on any level, the SOTA claim must be narrowed; if all are worse and separated, the claim is retained.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that SCULPT achieves state-of-the-art geometry on PartObjaverse at the part, semantic-group, and object levels. For that claim to be supported, the comparison set must include the methods the paper itself identifies as closest to SCULPT. Section 2.3 names UniPart and AutoPartGen as the nearest part-structured generators, yet neither appears in Table 1; the table is limited to Part123, OmniPart, and segmentation-plus-completion pipelines. The statement 'lowest CD and highest F1@.05' is therefore an assertion about a subset of the field, not about the state of the art. The risk is compounded by the small deltas: object-level CD is 0.0020 versus 0.0021 against TRELLIS.2+PartField, and no per-object variance, confidence interval, or paired test is reported. If either missing baseline outperforms SCULPT on any evaluation level under the same protocol, the headline claim is false as stated. This is a scope-of-evidence problem rather than an internal inconsistency, but it is the weakest load-bearing point of the empirical contribution.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces SCULPT, a recurrent subtractive decomposition method for part-aware 3D generation. Starting from a complete-object latent produced by TRELLIS.2 from a conditioning image, a joint split predictor iteratively predicts one extracted part together with the updated remainder in the shared object frame, using overlapping sparse supports, a composition loss, and inference-time support clipping. The rollout terminates when the remainder support is empty or when a fixed cap Kmax=24 is reached. Training uses part-annotated PartVerse-XL, and evaluation on PartObjaverse reports Chamfer distance and F1 at the part, semantic-group, and object levels, together with ablations and qualitative examples. The paper claims state-of-the-art geometry on PartObjaverse and presents the subtractive formulation as the main contribution.","tokens_in":18259,"tokens_out":7707,"duration_ms":74514,"significance":"The subtractive formulation is a genuinely different design point from additive generation and post-hoc segmentation, and the paper's ablations support the importance of full-stage adaptation, composition supervision, and inference-time clipping. If the empirical result holds over a complete baseline set, SCULPT would provide a practical way to obtain aligned, textured parts while preserving whole-object quality. Strengths include a clear methodological presentation, use of a strong pretrained backbone with zero-initialized control injection, and a three-level evaluation that separates component fidelity from assembly fidelity. The main weakness is that the state-of-the-art claim is currently supported only against a subset of the field, with small margins and no uncertainty quantification.","major_comments":[{"comment":"The related work identifies UniPart and AutoPartGen as the closest part-structured generation methods to SCULPT, yet neither appears in Table 1; the comparison is limited to Part123, OmniPart, and segmentation-plus-completion pipelines. Since the headline claim is state-of-the-art geometry on PartObjaverse, the absence of the two closest baselines leaves that claim unsupported. Please add these methods under the same evaluation protocol or explicitly restrict the claim to the evaluated baseline set.","section":"Section 2.3 / Table 1"},{"comment":"The margin over the strongest included baseline, TRELLIS.2+PartField, is small (part-level CD 0.0107 versus 0.0115; object-level CD 0.0020 versus 0.0021), and SCULPT is numerically worse on part-level and semantic-group F1@.1 (0.8858 versus 0.8897 and 0.8851 versus 0.8903). No variance, confidence intervals, or paired significance tests are reported, so the reader cannot tell whether the ordering is stable. Please report standard deviations or confidence intervals for all metrics, per-class results for all baselines, and a paired test over the 200 evaluation meshes.","section":"Section 4.2 / Table 1"},{"comment":"The Hungarian matching protocol for the part-level scores is underspecified for variable cardinailty: the paper does not state the cost matrix used for matching or how unmatched predicted and reference parts are penalized. The semantic-group procedure, in which each matched prediction inherits the semantic label of its reference part, can also conceal over-segmentation errors if extra predictions are matched incorrectly. Please specify the matching and scoring rules in enough detail that all methods are compared under exactly the same protocol.","section":"Section 4.1 / Metrics"},{"comment":"The method assumes that the structured latent and decoder can faithfully represent arbitrary sub-objects and overlapping interface shells, since parts and remainders are encoded and decoded independently in the shared object frame. This assumption is never verified. Please add a reconstruction experiment in which annotated part and remainder meshes are encoded, decoded, and compared against the original meshes with CD/F1 and qualitative inspection of the overlap boundaries. If reconstruction degrades for partial or overlapping supports, the central geometry claim needs qualification.","section":"Section 3.2 / 3.3"},{"comment":"The paper does not report how often the rollout hits the cap Kmax=24 on PartObjaverse or the distribution of predicted part counts. If the cap is frequently reached, the claim that the number of parts follows the rollout length is weakened, and the final retained remainder becomes a catch-all output. Please report the cap hit rate and predicted part-count statistics, and compare them with the ground-truth part-count distribution on the evaluation set.","section":"Section 3.5 / Algorithm 1"}],"minor_comments":[{"comment":"The abstract claims 'state-of-the-art geometry' without qualification, but Table 1 shows that SCULPT is not the best on part-level and semantic-group F1@.1; please qualify the claim to match the reported metrics.","section":"Abstract / Section 4.2"},{"comment":"The term 'O-V oxel' appears to be a typo or needs a consistent definition; the manuscript uses 'O-Voxel' elsewhere in the same section.","section":"Section 3.2"},{"comment":"Table 4 uses the headers 'F1-0.1' and 'F1-0.05' while Table 1 uses 'F1@.1' and 'F1@.05'; please unify the notation.","section":"Table 4"},{"comment":"The qualitative section does not identify the text-to-image model used to produce the fifth row of Figure 4; please state the model for reproducibility.","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"The core idea is sound and the paper is well written, but the empirical claim currently rests on an incomplete baseline set and unreported statistical uncertainty. If the authors cannot obtain UniPart and AutoPartGen under the same protocol, the conclusions should be scaled back to the evaluated baselines. The metric specification for Hungarian matching is also important for the paper's comparability and should be addressed in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The recurrent subtractive formulation is the real news here, and the paper does a good job showing it works. The idea of a fixed-signature split that emits one part plus an explicit remainder, with termination learned from an empty remainder, is a clean departure from both segmentation-then-label and additive part synthesis. Jointly denoising the part and remainder on the union of their sparse supports, letting boundary voxels overlap, is a sensible answer to the usual gap/interpenetration problem. The ablations are well designed and informative: voxel-only support prediction collapses, composition supervision matters, and inference-time clipping matters. The training data is substantial, the evaluation uses an external benchmark with a stated dedup protocol, and the paper is mostly clear about what was done.\n\nBut the \"state-of-the-art\" sentence in the abstract is stronger than the evidence. UniPart and AutoPartGen, which the paper itself names as the closest part-structured generators, are absent from Table 1. Against the strongest included baseline, TRELLIS.2+PartField, the advantage is modest: part-level CD 0.0107 vs 0.0115, object-level CD 0.0020 vs 0.0021, and SCULPT actually loses on part- and semantic-group-level F1@0.1. There is no per-object variance, confidence interval, or paired test anywhere. If either missing baseline outperforms SCULPT on any of the three levels under the same protocol, the headline claim is simply false. This is a scope-of-evidence problem, not an internal inconsistency, but it is the weakest load-bearing point.\n\nOther soft spots are minor by comparison. The occupancy threshold tau_occ is never given a value, and it directly controls part counts and termination. No code is released, so the numbers cannot be independently checked. The qualitative set is tiny: four dataset images, one text-to-image input, and one photo. The paper also never directly verifies the assumption that TRELLIS.2's structured latent can faithfully represent arbitrary sub-objects, but the strong empirical results and the voxel-only ablation make that worry less pressing; it is a fair referee question, not a flaw that sinks the paper.\n\nI do not see a circularity problem here. The evaluation is against an external benchmark with independent annotations and a SHA-based dedup step, and the heavy use of TRELLIS.2 as a backbone is legitimate when the method adapts it and compares against it.\n\nWho is this for: anyone working on part-aware 3D generation or editable 3D assets. I would bring it to a reading group and I would cite it for the formulation. It deserves a serious referee, but the referee should push hard on the missing baselines and the unquantified margins. Send it to review; require UniPart and AutoPartGen in the comparison table, variance reporting, and the tau_occ value before accepting.","headline":"The recurrent part-remainder formulation is genuinely new and well demonstrated, but the state-of-the-art claim overreaches: the closest baselines are missing and the margins over the strongest included one are small and unquantified.","tokens_in":18796,"tokens_out":2098,"would_cite":true,"duration_ms":22907,"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":"Part-aware 3D generation can be done by subtractive composition, repeatedly splitting a part and the remainder from a whole-object latent, and the paper reports the best geometry on PartObjaverse.","keywords":["3D part generation","subtractive composition","part-aware generation","structured 3D latent","rectified flow transformers","recurrent decomposition","image-to-3D","part segmentation and assembly"],"falsifier":"Take a held-out object, encode its complete mesh and also encode one of its part meshes together with the complementary remainder mesh in the backbone's structured latent, decode each independently, and merge the two outputs; if the merged Chamfer distance to the original complete mesh is much larger than the reported object-level value of about 0.0020, or if visible gaps appear at the shared interface, the subtractive-composition assumption fails. A cheaper check is to compare the geometry and material colors on the overlapping interface-shell voxels predicted from the part side and from the remainder side of a split and measure their disagreement.","tokens_in":17768,"feed_emoji":"🧩","tokens_out":13430,"duration_ms":131907,"temperature":0.7,"pith_summary":"SCULPT claims that part-aware 3D generation is best done subtractively: instead of labeling an already generated shape or synthesizing parts and then reconciling them, the model starts from a complete object in a generative 3D latent space and repeatedly predicts one extracted part together with the object that remains. Each split is a fixed-signature operation, so the number of parts is determined by how many splits the rollout performs, up to a safety cap, rather than by a fixed number of output slots. The part and remainder are denoised together on overlapping supports in a shared object frame, which is designed to keep contact boundaries, geometry, and materials coherent. On the part-annotated PartObjaverse benchmark the paper reports the lowest Chamfer distance and best F1 at the part, semantic-group, and object levels, meaning the parts are individually accurate and the assembled whole stays faithful. If the claim holds, part editing, material assignment, and reassembly become outputs of generation rather than post-hoc fixes.","feed_headline":"Subtractive 3D part generation beats additive baselines","feed_subtitle":"The model repeatedly splits a part off the whole object, scoring best on part-annotated 3D geometry.","key_machinery":"The load-bearing object is the joint split predictor, a recurrent two-output module that turns the current 3D state into one extracted part and the updated remainder. It operates on the structured latent of the backbone generator, which has sparse-structure, geometry, and material stages in a shared $[-1,1]^3$ object frame; part and remainder are packed onto the union of their sparse supports, denoised jointly by rectified-flow transformer blocks initialized from the pretrained whole-object generator, and controlled by a remainder-injection branch whose residual projections start at zero. A composition loss compares the differentiable union of the predicted part and remainder occupancies with the incoming remainder's support, allowing an overlapping interface shell instead of a disjoint voxel partition, and inference-time clipping keeps every predicted support inside the current object. This machinery is what lets one fixed-signature operation express an object-dependent number of parts through rollout length, and it is what the paper credits for the part-level and object-level geometry scores.","core_discovery":"The paper's central discovery is that a variable-cardinality part decomposition can be produced by composing a single two-output prediction rather than by predicting the whole part set at once. Given a complete-object latent from a pretrained image-to-3D generator, a joint split predictor emits the next part and the updated remainder in a coupled denoising trajectory; the predictor is then applied again to the remainder, and the rollout stops when the remainder's support is empty or a cap of 24 splits is reached. Because the part and remainder stay in the coordinate frame of the original object and may share overlapping boundary voxels, the boundary between them is decided during generation instead of imposed by a voxel-face partition. The paper reports that on PartObjaverse this yields the best Chamfer distance and best F1 at threshold 0.05 at the part, semantic-group, and object levels, with object-level Chamfer distance 0.0020 and F1 at threshold 0.05 of 0.9212, and that the approach also produces textured decompositions from text-to-image inputs and a real photograph. The authors would summarize the result as: subtractive composition turns the difficulty of modeling how many parts an object has into the easier question of when the remaining object is empty.","pith_inferences":["A natural extension the paper does not pursue is porting the split interface to other whole-object generators that have a sparse structured latent and a decoder, since the formulation itself does not depend on the specific backbone.","The deterministic centroid-based extraction order is a convention rather than a requirement; a learned or user-selected split order could expose hierarchical or part-graph structure using the same machinery.","The overlapping interface shell makes a direct boundary-coherence test possible: comparing geometry and material on voxels predicted from both sides of a split would quantify how much of the gain comes from joint denoising.","The paper's cap of 24 splits is a practical bound, and the demonstrated recursive application to already-extracted parts suggests that a calibrated stopping rule could make decomposition depth-adaptive rather than capped."],"forward_implications":["Part structure becomes a native output of generation: the extracted parts carry geometry, material, and texture and can be edited or reassigned without a separate segmentation step.","Variable part counts emerge from rollout termination, so the model is not limited by a predefined number of slots or an image-space layout plan.","Because every part is decoded in the shared object frame, the assembled union needs no registration, rescaling, or snapping, which preserves the complete-object silhouette.","The benchmark numbers imply that fine-grained decomposition does not trade away whole-object fidelity: the best object-level scores are obtained by the same model that produces the best parts.","The same rollout extends beyond the benchmark to text-to-image inputs and a real photograph, suggesting the subtractive procedure transfers to casual images."],"supporting_citations":[{"why":"Supplies the structured sparse-voxel latent with geometry and material stages, the pretrained checkpoints the split predictor is initialized from, and the decoder that produces meshes in the shared object frame.","marker":"(Xiang et al., 2026)"},{"why":"Introduced the structured 3D latent formulation that the whole-object generation prior builds on, giving the subtractive process its complete-object starting point.","marker":"(Xiang et al., 2025)"},{"why":"Provides PartVerse-XL, the part-annotated assets whose complete mesh plus ordered part meshes yield the per-split part–remainder supervision tuples.","marker":"(Ding et al., 2026)"},{"why":"Released the PartObjaverse benchmark with instance and semantic part annotations used for the three-level geometry evaluation.","marker":"(Yang et al., 2024d)"},{"why":"OmniPart is the strongest direct part-generation baseline; the paper's part- and object-level improvements over it anchor the state-of-the-art claim.","marker":"(Yang et al., 2025)"},{"why":"PartField provides the post-hoc decomposition baseline that, combined with the whole-object generator, is SCULPT's closest competitor.","marker":"(Liu et al., 2025)"},{"why":"AutoPartGen is the closest autoregressive part-generation relative; the paper contrasts its lack of an explicit remainder state with SCULPT's carried remainder.","marker":"(Chen et al., 2025b)"},{"why":"ControlNet-style zero-initialized residual injection is the mechanism the remainder-control branch uses to condition the pretrained denoising blocks on the current 3D state.","marker":"(Zhang et al., 2023)"}],"fun_headline_variants":["Split to sculpt: adaptive 3D part generation","SCULPT: cutting 3D parts straight from images","Iterative splitting yields top 3D part geometry","One split at a time beats additive 3D assembly"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the pretrained generator's internal 3D representation can faithfully represent partial sub-objects and overlapping contact regions, so a part and a remainder decoded separately and merged reproduce the complete object.","fun_headline_variants_meta":{"raw":{"variants":["Split to sculpt: adaptive 3D part generation","SCULPT: cutting 3D parts straight from images","Iterative splitting yields top 3D part geometry","One split at a time beats additive 3D assembly"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000595,"raw_usage":{"total_tokens":2860,"prompt_tokens":1092,"completion_tokens":1768,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":708,"completion_tokens_details":{"reasoning_tokens":1702}},"tokens_in":708,"tokens_out":1768,"duration_ms":14927,"temperature":1.0,"reasoning_tokens":1702,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:34:03.850831+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a held-out object, encode its complete mesh and also encode one of its part meshes together with the complementary remainder mesh in the backbone's structured latent, decode each independently, and merge the two outputs; if the merged Chamfer distance to the original complete mesh is much larger than the reported object-level value of about 0.0020, or if visible gaps appear at the shared interface, the subtractive-composition assumption fails. A cheaper check is to compare the geometry and material colors on the overlapping interface-shell voxels predicted from the part side and from the remainder side of a split and measure their disagreement.","supporting_citations":[{"cited_title":"Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =","cited_arxiv_id":null,"evidence_quote":"ControlNet-style zero-initialized residual injection is the mechanism the remainder-control branch uses to condition the pretrained denoising blocks on the current 3D state."}],"review_version":1}