acceptodds
Withdrawn from ICLR 2027

3D Primitive Assembly from Generative Anchors

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

10 views

Abstract

Many 3D objects can be approximated as assemblies of simple primitives, turning a surface into an editable structure of parts. Recovering that assembly from an observed 3D model requires both a global account of which parts exist and precise localization on the input. An image follows the same conversion after primitive-style abstraction and surface reconstruction. Directly generating primitive parameters must invent unobserved interiors, and overlapping solids often admit several valid programs. A detailed mesh defers the type choice, but spends many tokens on coordinates that still do not partition the observed surface. Point features preserve geometric detail, yet do not decide how many regions to keep; small exposed attachments are easily dropped. We recover primitive assemblies from 3D generative anchors. Conditioned on ShapeVAE, an autoregressive model generates a low-density proxy mesh disconnected along primitive contacts. These anchors are not the delivered geometry: connected components organize their hidden states into region queries, a point transformer predicts dense assignment and primitive type on the processed input surface, and analytic fitting restores editable bodies. Geometric precision is therefore carried by the dense surface, while expensive autoregression only produces token-efficient query summaries. The pipeline yields a variable-part assembly without a primitive-parameter language and without a prescribed part count.

Primary area:
generative models
Keywords:
3D generative models, Primitive assembly, Autoregressive mesh generation, 3D part segmentation, Point transformers, Geometry processing

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_gLEvOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a method to convert an input mesh into a set of simple primitives. The paper proposes to use sparse anchors mesh component as a an intermediate representation to ease the fitting of primitive while dealing with hidden/intersecting regions.

Critical strengths
  1. The insight that a form geometry guided segmentation - here performed trough the proposed anchors - helps more robustly fit primitives
  2. The final result, as seen from figure 9 for instance, seem significantly better than the base line they compare to
Critical weaknesses
  1. The manuscript is globally more a description of a system than a well grounded research paper. Most of its content are statement rather than proposition stemming from an analysis of the problem
  2. It remains unclear whether alternative, simpler segments would not lead to surface regions that are good candidate for the fitting stage
  3. Little to no care is attached to the geometric analysis: for instance, intrinsic (and even extrinsic) symmetries are ignored (as can be seen in a number of examples, where exact copies of geometries structures get decomposed differently
  4. I cannot assess how one can control the level of detail of the resulting assembly
  5. The text is hard to read: short, very compressed sentence and out of the blue statement, as too often produced by LLM in a performative ton
  6. The papers consider only very recent prior art, and does not discuss earlier learning-based approach, nor pure geometric methods e.g., GlobFit
  7. Most example reported consists in input mesh that were clearly themselves made from some primitive-based modeling process. For the few that do not clearly come from such a process, such as the horse or the bull on figure 14, the dominance of the proposed method against the baselines is less clear.
Questions for authors
  1. How does the method compare to pure geometric analysis e.g., hough transform like
  2. Should the scope of the claim be toned down, toward categories of input shape that are essentially made of simple primitives in the first place?
  3. how does the method perform on incomplete shapes (e.g., scans) ?
  4. how robust is the method to the initial sampling both in term of density and uniformity?
Reviewer_h946OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces PAGA, a method for converting a detailed 3D mesh into a compact, editable assembly of geometric primitives through two main stages: coarse part discovery and fine-grained primitive fitting. In the coarse stage, an autoregressive transformer initialized from MeshAnything V2 takes sampled surface points and normals as input and generates sparse geometric anchors whose connected components represent candidate part regions. Features associated with each component are pooled into queries, which assign points on the original surface to the proposed regions. In the fine-grained stage, a learned classifier predicts primitive-family preferences, guiding numerical fitting of primitive types (e.g. boxes, prisms, cylinders, cones, frusta, ellipsoids, and tori) and parameters to the assigned surface points. Experiments on 2,000 near-primitive surfaces and 2,000 Objaverse meshes show lower surface reconstruction error than the evaluated baselines, with fewer primitives than PrimitiveAnything and Marching Primitives.

Critical strengths

Promising idea and results. Combining generative part discovery with surface-based fitting is well motivated and presented results are encouraging. PAGA achieves lower surface reconstruction error than the evaluated baselines and uses fewer primitives than PrimitiveAnything and Marching Primitives on both test sets. However, provided details and relevant baselines are missing.

Critical weaknesses
  • Missing crucial details and unclear exposition. The method section omits important details needed to understand and reproduce the pipeline: (a) the ordering used to serialize sparse proxy meshes for autoregressive supervision, although the next-token objective depends on this order; (b) the primitive initialization and fitting procedure, including its objective, candidate-selection rules, and acceptance criteria (L293–311); and (c) the provenance and construction of training data and near-primitive evaluation set, including how the ground truth primitive decompositions used for supervision are obtained.
  • Missing comparison with a strong baseline. PAGA compares against Marching Primitives and PrimitiveAnything but omits SuperFit. SuperFit reports better reconstruction quality than these methods on its own benchmarks, making it a relevant baseline to evaluate on PAGA’s datasets.
Questions for authors
  1. How are sparse proxy meshes serialized for AR supervision? Is the slot ordering inherited unchanged from MeshAnything V2?
  2. How are different primitives initialized and fitted? More details should be added in 293-311, the current description is very high-level.
  3. Could you provide more details on the source of the training dataset and the “near-primitive surfaces” evaluation set? How are the GT primitive decompositions obtained?
  4. Would it be possible to add SuperFit among the proposed baselines?
  5. Could you highlight the best result for each metric in bold in the quantitative tables to make comparisons easier to read?
  6. Could you explain MeshAnything V2’s substantially higher reconstruction error in Table 2, including the main failure modes and the preprocessing and inference settings used for this baseline?
Reviewer_NYCAOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes a 3D mesh to primitive-based assembly conversion framework (PADA). The main idea of this paper is to separate part discovery from precise geometric fitting. An autoregressive model generates sparse, contact-separated anchors, whose connected components provide queries for identifying corresponding regions on the original dense surface.

Critical strengths

1. Compelling approach of separating structural discovery and precise geometry estimation. The main idea of using a generative model to propose part candidates rather than directly predict the final geometry is intuitive and practically solid. Authors propose separation of sparse anchor prediction and geometric fitting, which can mitigate the limitation of previous approaches.

2. Sound design choice and supporting experimental results. The main components are aligned with the overall idea. Each methodological component such as contact-separated sparse supervision, fitting that permits interior completion are technically well designed and the promising experimental result with qualitative comparison shows the benefits of the proposed approach and technical design.

Critical weaknesses

1. Insufficient evaluation design.

  • Insufficient quantitative ablations of the main components - Although the full framework performs well, the individual contributions of each components such as learned grounding, interior-aware fitting, and consolidation are not sufficiently examined. Without quantitative comparisons, it is difficult to determine which design choices account for the improvements and how much each contributes.

  • Wrong user study protocol - The paper reports that 10 participants from the authors' lab evaluated outputs with method names hidden, and that all preferred PAGA in terms of orderliness and visual appeal. However, the details of human evaluation are missing and the participant pool is too small. Such uncontrolled experimental result looks not appropriate as a user study. If authors are planning to claim PADA is superior in human perception perspective, the details of survey and validity of that comparison should be provided.

2. Limited recovery from structural errors in the initial anchor hypotheses. PAGA can refine anchor geometry through grounding and fitting, but its part decomposition remains constrained by the generated connectivity. Consolidation can mitigate over segmentation, but no explicit mechanism is described for recovering missing or incorrectly merged parts, and the mesh-based fallback does not necessarily restore a compact decomposition. The paper does not quantify how frequently these structural errors occur or their impact on the compactness and organization of the final assembly.

Questions for authors

1. Component ablations. Could you quantify the contributions of learned grounding, interior exemption, and consolidation using the existing models? Comparisons against nearest-anchor assignment, fitting without interior exemption, and pre-consolidation outputs, reporting surface fidelity and primitive count on a common test subset, would help isolate their effects.

2. User evaluation. Could you clarify the number and selection of evaluated objects, the presentation protocol, and whether preferences were recorded per object or overall? Please also distinguish evidence of visual preference from evidence of practical editability.

3. Structural error recovery. Does the pipeline use fitting residuals to recover missing or incorrectly merged part hypotheses? Please analyze the frequency of these errors and fallback usage, along with their effects on reconstruction quality and primitive count, using existing test outputs.

Overall, I am currently leaning toward acceptance, but I strongly recommend that the authors strengthen the experimental evidence through the additional analyses outlined above. Please also review the manuscript’s formatting and ensure that it follows the official conference template and formatting guidelines.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.