{"id":"4d0172e2-6a6b-4ee4-a549-836bf2793c3d","arxiv_id":"2507.06165","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A two-stage pipeline plans 3D part boxes from an image plus 2D masks, then synthesizes coherent part-level geometry by adapting the pretrained TRELLIS generator.","lead":"OmniPart turns a single image into a 3D model made of separate, editable parts by first predicting part boxes and then generating all parts together. A generalist reader should care because editable 3D parts are what make games, animation, and product design workflows practical.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 2's part-level CD/F1 metrics lack any stated matching protocol between predicted and ground-truth parts, so the thin SOTA margin (F1-0.1 0.74 vs 0.69) is not independently checkable.","rationale":"The paper's central claim is that OmniPart achieves state-of-the-art part-aware 3D generation. That claim rests primarily on Table 2, where the margins over the strongest baseline are thin (CD 0.18 vs 0.19; F1-0.1 0.74 vs 0.69). The most load-bearing assumption is therefore that the part-level metrics are computed in a way that is comparable across methods and independently verifiable. The manuscript defines a matching procedure for bounding boxes in Table 1, but for Table 2 it gives no correspondence or association protocol for part-level CD/F1. Since the model generates a variable number of parts and the paper explicitly disclaims one-to-one correspondence, the metric is underspecified. Without a fixed protocol, one cannot tell whether the reported improvement reflects the method or an unspecified matching choice. This is internal to the paper's own evaluation, whereas the train/test split concern identified by the reader is also valid but concerns external validity. I therefore flag the missing part-matching definition as the single most load-bearing concern. The two-stage design is coherent, the voxel-discarding mechanism is plausible, and the ablations are informative, so the paper remains a credible conditional contribution if the authors can specify and release the evaluation protocol. Otherwise the state-of-the-art claim is unverifiable from the manuscript as written.","tokens_in":15997,"tokens_out":4615,"duration_ms":47505,"concrete_test":"Ask the authors to release the predicted part meshes for all 300 test objects and the exact algorithm used to compute part-level CD and F1 for Table 2. Independently recompute these metrics with a fixed protocol, e.g., Hungarian matching between predicted and ground-truth part point clouds using Chamfer distance as the cost, and check whether OmniPart's F1-0.1 margin over TRELLIS+PartField (0.74 vs 0.69) persists. If it does not persist under this fixed protocol, the state-of-the-art claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.2 reports part-level CD and F1 as the primary evidence for state-of-the-art part-aware generation (Table 2: OmniPart CD 0.18 vs TRELLIS+PartField 0.19; F1-0.1 0.74 vs 0.69). However, the manuscript never specifies how predicted parts are matched to ground-truth parts when computing these part-level metrics. This matters because Section 4.1 explicitly states that the generated bounding boxes have 'no one-to-one correspondence' with ground truth, and the model outputs a variable number of parts. Without a stated matching algorithm (e.g., Hungarian assignment on part meshes, nearest-part point assignment, or union-of-parts comparison), part-level CD/F1 can vary substantially; a method that over-segments or merges parts could alter the scores. The reported margin over TRELLIS+PartField is small, so the ranking could invert under a different, equally reasonable matching protocol. This is an omitted definition at the core of the empirical claim, not a tuning detail.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"OmniPart proposes a two-stage pipeline for part-aware 3D generation from a single image and user-specified 2D masks. The first stage autoregressively predicts 3D part bounding boxes conditioned on DINOv2 image features, per-pixel mask embeddings, and TRELLIS coarse voxels, trained with an added coverage loss. The second stage fine-tunes TRELLIS's rectified flow model to denoise part latents jointly with whole-object context, using part position embeddings and an extra voxel-validity channel to discard extraneous voxels. Experiments compare bounding-box planning and part-level/whole-object generation against several baselines and report state-of-the-art numbers, with qualitative demonstrations of mask control, multi-granularity generation, material editing, and remeshing.","tokens_in":16129,"tokens_out":7132,"duration_ms":74616,"significance":"If the reported results generalize, OmniPart is a practically useful contribution: the two-stage decomposition is clean, the mask-conditioned autoregressive planner offers an intuitive controllability mechanism, and the efficiency gain (0.75 min vs. 5–15 min in Table 3) is substantial. The paper gives appropriate credit to the coverage-loss ablation in Table 1, and the voxel-discarding mechanism is a sensible treatment of boundary noise in coarse voxel representations. However, the central \"state-of-the-art\" claim rests on an evaluation whose train/test separation and metric definitions are not fully specified; until those are clarified, the numerical advantage over TRELLIS+PartField and PartGen should be treated as preliminary.","major_comments":[{"comment":"The 300-object test set is sampled from the dataset described in Section 3.4, but no train/test split is specified, and Section 3.4 states that the autoregressive bounding-box model is trained on all 180K annotated shapes. If the test objects (or their part annotations) appear in the training pool, the quantitative margins in Tables 1 and 2—e.g., part-level F1-0.1 0.74 vs. 0.69 in Table 2—could be inflated by memorization rather than generalization. Please state explicitly how the 300 test objects were held out from both the 180K planning training set and the 15K fine-tuning set, and report metrics on a held-out subset.","section":"§4 Evaluation Protocol; §3.4"},{"comment":"Part-level Chamfer Distance and F1 require a defined correspondence between predicted and ground-truth parts, but the manuscript never specifies the matching procedure. Section 4.1 explicitly notes that predicted bounding boxes have no one-to-one correspondence with ground truth, and Section 3.2 emphasizes non-one-to-one mask conditioning. Without a stated matching algorithm (e.g., Hungarian matching on part meshes, nearest-part assignment, or union-of-parts evaluation), the part-level scores are not independently checkable; since the margin over TRELLIS+PartField is small (CD 0.18 vs. 0.19, F1-0.1 0.74 vs. 0.69), the ranking could flip under a different reasonable protocol. Please define the matching protocol and, ideally, report results under an alternative protocol as a robustness check.","section":"§4.2, Table 2"},{"comment":"The training-data section describes how part bounding boxes and voxel latents are constructed, but it never states how the 2D part masks that condition the autoregressive planner are obtained for the 180K training shapes. Since 2D mask conditioning is a core contribution and is used to control part granularity, the paper should specify whether masks are rendered from 3D part annotations, produced by SAM, or synthesized, and how the non-one-to-one correspondence between masks and 3D boxes is realized in training. Without this, the main method cannot be reproduced.","section":"§3.4"},{"comment":"For a state-of-the-art claim, the comparison set is incomplete: PASTA is discussed in Related Work, and PartCrafter and Dual Volume Packing are cited, but none appears in Table 2. Please either add these methods to the quantitative comparison or explicitly state the reasons for exclusion (e.g., input modalities or code availability) and qualify the state-of-the-art claim accordingly.","section":"§2.2, §4.2"}],"minor_comments":[{"comment":"The caption states F1-score thresholds CD<0.1 and CD<0.05, but the column headers read F1-0.1 and F1-0.5; please resolve the mismatch (likely F1-0.05) so the reported numbers are interpretable.","section":"Table 2 caption"},{"comment":"Tables 1 and 2 report single runs without error bars or significance tests; given the thin margins in Table 2, please report standard deviations over multiple seeds or bootstrap confidence intervals.","section":"§4, Tables 1–2"},{"comment":"Reporting the best score over four rotations is a favorable-evaluation choice; please also report the mean over rotations so readers can assess orientation sensitivity.","section":"§4.2, Table 2"},{"comment":"For the voxel-level metrics, the matching procedure is described only as \"proximate predicted counterpart\"; please specify the matching rule (e.g., Hungarian assignment on bounding-box overlap) for Table 1 as well.","section":"§4.1, Table 1"},{"comment":"The coverage loss is defined over min/max token positions, but it is not stated whether coordinates are normalized globally or per object; adding this detail would aid reproducibility.","section":"§3.2, Eq. (2)"}],"recommendation":"major_revision","confidential_remarks":"The paper is timely and the architecture is plausible; my recommendation of major_revision is driven by the evaluation protocol, not by the method itself. I would ask the editor to require a clear train/test split and exact part-matching definitions before considering a revised version. The missing baselines are partly explained by concurrent work, but PASTA predates this submission and should be addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: OmniPart is a coherent two-stage system for part-aware 3D generation from a single image, and the core architectural idea — mask-conditioned autoregressive box planning followed by a TRELLIS fine-tuned rectified flow that generates all parts jointly — looks sound. The paper's main weakness is evaluational: the part-level metrics that back the SOTA claim rest on an unspecified matching procedure, the test set is sampled from the same pool as training without a stated split, and several directly comparable baselines are cited but not run.\n\nWhat's new: the specific combination plus the coverage loss and the voxel-validity channel. The coverage loss is simple but its ablation shows it matters; the voxel-discarding mechanism is a sensible fix for the noise-box problem and is well motivated. The efficiency numbers are striking.\n\nWhat's soft: the stress-test concern is real. Table 1 says predicted boxes have no one-to-one correspondence with GT, and then Table 2 reports part-level CD/F1 without ever saying how predicted parts are matched to GT parts. With the margin over TRELLIS+PartField at 0.18 vs 0.19 CD and 0.74 vs 0.69 F1, a different reasonable matching rule could flip the ranking. The authors need to specify the matching protocol or report a union-of-parts / Hungarian assignment.\n\nAlso, the 300-object test set is sampled from the dataset used for training, and no split is stated. That's an evaluation hygiene problem, not proof of a leak, but it makes the numbers hard to trust. And PASTA, PartCrafter, Dual Volume Packing are all cited in related work but absent from Table 2; that's an incomplete benchmark for a SOTA claim.\n\nOn the positive side, the ablations are informative, the paper is clearly written, and the design choices are well motivated. The circularity burden is low.\n\nVerdict: the architecture is worth building on, but the empirical claims need more work before I'd trust them. This deserves peer review — a serious referee could push for the missing experiments and a clearer evaluation protocol.","headline":"Solid architecture, shaky headline: the two-stage design is worth attention, but the SOTA claim rests on under-specified part-level metrics and an incomplete benchmark.","tokens_in":16777,"tokens_out":2497,"would_cite":true,"duration_ms":24941,"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":"OmniPart generates 3D objects as editable parts from a single image by first planning part bounding boxes and then synthesizing all parts together.","keywords":["3D generation","part-aware generation","single-image 3D reconstruction","autoregressive structure planning","rectified flow","sparse voxel latents","2D mask conditioning","compositional 3D assets"],"falsifier":"Re-run the Table 2 comparison on a test set whose categories are held out of the 180K training collection, and check whether the part-level Chamfer distance and F1 margins over the closest segmentation-plus-holistic baseline persist; if they collapse to 0.19 and 0.69, the claimed advantage is a training-set artifact.","tokens_in":15664,"feed_emoji":"🧊","tokens_out":6959,"duration_ms":62866,"temperature":0.7,"pith_summary":"The paper introduces OmniPart, a framework that turns a single photograph plus rough 2D part masks into a 3D object built from distinct, editable parts. It argues that the right way to get part-aware 3D generation is to separate high-level structure planning from detailed geometry synthesis. An autoregressive module predicts a variable-length list of 3D part bounding boxes guided by 2D masks, and a rectified-flow model, adapted from a pre-trained holistic generator, denoises all part latents jointly so parts stay consistent and align with the planned layout. Experiments report state-of-the-art part-level accuracy, with Chamfer distance 0.18 and F1 0.74 at threshold 0.1, while the whole merged object also improves over generating the object monolithically. If this holds, image-to-3D tools could offer manipulable, animatable assets rather than single fused shapes.","feed_headline":"Editable 3D parts from one image, with masks steering the split","feed_subtitle":"A two-stage pipeline plans part bounding boxes, then generates all parts together, beating prior part-aware methods","key_machinery":"The load-bearing mechanism is a two-stage decoupling coupled by sparse voxel coordinates. Stage one is an autoregressive bounding-box planner, whose sequence tokens are conditioned on masked image features and whole-object voxel tokens, with a part coverage loss that penalizes boxes too small to contain their part. Stage two is a spatially-conditioned rectified-flow part synthesizer: noisy latents for the whole shape and for each part are packed into a transformer, distinguished by part position embeddings with a shared index for the whole object and a unique index per part, and a validity channel per voxel decides which boundary voxels to discard. The planned boxes bridge the stages by giving each part a spatial region in the same voxel grid the pretrained generator uses.","core_discovery":"The central claim is that part-aware 3D generation can be decomposed into two synergistic stages that resolve the tension between semantic decoupling and structural cohesion. OmniPart first plans the layout: a transformer-based autoregressive model emits an ordered sequence of axis-aligned 3D bounding boxes, sorted z-y-x, conditioned on masked image features and coarse whole-shape voxel tokens; a part coverage loss forces each predicted box to enclose its true part generously. The second stage generates all parts at once: a rectified-flow denoiser, fine-tuned from a holistic pretrained generator, operates on sparse voxels inside each predicted box plus whole-object context, with per-part position embeddings and a learned voxel-validity channel that discards boundary voxels assigned to the wrong part. The result, by the paper's quantitative evaluation, is the best part-level geometry among compared methods while preserving whole-object quality, and the framework supports variable part granularity, 2D-mask-controlled decomposition, material editing, and animation.","pith_inferences":["An unstated consequence is that the 2D-mask interface could generalize beyond class-agnostic segmentation: if masks are replaced by text prompts or point clicks, the planner might inherit those controls without retraining the synthesizer.","A likely limit not stressed in the paper is that axis-aligned bounding boxes are a coarse container for curved or interlocking parts; the stated limitation about noisy boundary voxels implies that concave or deeply nested parts may be the first failure mode on out-of-distribution shapes.","A testable extension would be measuring part-level editability directly, for example replacing one generated part's material or pose and checking whether the unchanged parts remain geometrically compatible, since the paper demonstrates applications qualitatively but does not quantify editing robustness.","The reported speed gain suggests a practical deployment path: a planner plus a fine-tuned pretrained generator could serve part-aware generation in interactive tools, but only if the train/test separation in the 180K-shape dataset is confirmed to be category-disjoint."],"forward_implications":["Users can steer decomposition granularity by choosing coarser or finer 2D masks; the same image can yield a single merged object or many separately addressable parts.","Parts are generated simultaneously rather than one-by-one, so the pipeline runs in roughly 0.75 minutes per object, compared with 5 to 15 minutes for multi-view reconstruction baselines.","Because each part has independent geometry and texture while remaining globally consistent, downstream tasks such as material swapping, remeshing, and animation operate on clean part boundaries without post-hoc segmentation.","Merging the generated parts preserves or slightly improves whole-object fidelity (Chamfer distance 0.07, F1 0.93), suggesting that imposing explicit structure does not cost overall shape quality."],"supporting_citations":[{"why":"Supplies the structured sparse-voxel latent representation and the pre-trained holistic rectified-flow generator that the synthesis stage fine-tunes.","marker":"[Xiang et al. 2024]"},{"why":"Provides the class-agnostic 2D segmentation model used to obtain flexible part masks that condition the structure planner.","marker":"[Kirillov et al. 2023]"},{"why":"Serves as the zero-shot point-cloud segmentation baseline whose predicted part masks are compared with the planner and used in the TRELLIS+PartField baselines.","marker":"[Liu et al. 2025a]"},{"why":"Provides the completion-based part-conversion baseline (TRELLIS+PartField+HoloPart) that OmniPart is compared against at part level.","marker":"[Yang et al. 2025b]"},{"why":"Supplies the multi-view part reconstruction baseline Part123 used in the part-level and efficiency comparisons.","marker":"[Liu et al. 2024a]"},{"why":"Supplies the multi-view part generation baseline PartGen used in the part-level and efficiency comparisons.","marker":"[Chen et al. 2024b]"},{"why":"Provides the 3DShape2VecSet encoder that converts whole-object voxels into tokens conditioning the autoregressive planner.","marker":"[Zhang et al. 2023]"},{"why":"Defines the conditional flow matching objective used to train the spatially-conditioned part synthesis rectified flow model.","marker":"[Lipman et al. 2024]"},{"why":"Supplies the DINOv2 visual features used both to condition the planner on the input image and to build part latents from rendered views.","marker":"[Oquab et al. 2023]"}],"fun_headline_variants":["Part-aware 3D generation: plan boxes, then generate all parts at once","Masks control 3D part layout; flow model synthesizes parts together","Two-stage OmniPart: semantic decoupling for editable 3D assets","From one image to editable 3D parts, with mask-guided decomposition"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results depend on the 300 test objects being representative and unseen during training, which the paper does not explicitly verify.","fun_headline_variants_meta":{"raw":{"variants":["Part-aware 3D generation: plan boxes, then generate all parts at once","Masks control 3D part layout; flow model synthesizes parts together","Two-stage OmniPart: semantic decoupling for editable 3D assets","From one image to editable 3D parts, with mask-guided decomposition"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000863,"raw_usage":{"total_tokens":3744,"prompt_tokens":946,"completion_tokens":2798,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":2715}},"tokens_in":562,"tokens_out":2798,"duration_ms":21755,"temperature":1.0,"reasoning_tokens":2715,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:09:14.464291+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the Table 2 comparison on a test set whose categories are held out of the 180K training collection, and check whether the part-level Chamfer distance and F1 margins over the closest segmentation-plus-holistic baseline persist; if they collapse to 0.19 and 0.69, the claimed advantage is a training-set artifact.","supporting_citations":[],"review_version":1}