acceptodds
Withdrawn from ICLR 2027

CADPalette: Geometric Primitive Segmentation of CAD Models via Palette-Coded Structured 3D Latents

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

12 views

Abstract

Geometric primitive segmentation of CAD surfaces assigns each face a geometric type while preserving the boundaries of individual surface components. This task is difficult when discretization varies widely, because evidence learned from one mesh domain can change substantially across native CAD tessellations, regularized remeshes, and scan reconstructions, where altered sampling, resolution, and noise can destabilize boundaries. We introduce CADPalette, an instance-aware geometric primitive segmentation framework that carries primitive types and component membership along one structured 3D generative trajectory. Each surface is converted to Shape O-Voxel and Shape SLAT conditions, while a Palette O-Voxel stores fixed primitive bits and within-shape randomized instance colors in a shared carrier. A fine-tuned flow predicts the corresponding Palette SLAT, and one frozen material SC-VAE decodes both fields for face-level primitive-type and instance readout. Explicit RGB separation, a frozen SC-VAE response-interpolator gate, and latent relation supervision constrain the instance field from target construction through decoding. In the common validation protocol, CADPalette surpasses the baselines by 13.3% points in primitive-type Prim-mIoU and 14.9% points in primitive-type Prim-wIoU, while also surpassing the baselines by 21.5% points in Inst-mIoU, 7.3% points in ARI, and 5.3% points in Pair-F1. Extensive experiments show that CADPalette achieves stronger geometric primitive segmentation performance and better generalization than the existing methods.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
CAD reverse engineering, CAD surface understanding, geometric primitive recognition, instance segmentation, flow matching

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_YmfuOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper presents CADPalette, which uses the material latent of a pretrained 3D generative model to predict both geometric primitive types and surface-instance partitions for CAD meshes. Primitive types are encoded as fixed bits and instances as random colors. A fine-tuned material flow produces the joint representation, and color screening together with latent relation supervision reduces confusion between adjacent instances. The authors run comparisons and ablations on a CAD dataset they constructed, and show qualitative results on scan-reconstructed meshes.

Critical strengths

On the reported CAD validation set, the method is clearly better than the selected baselines at both geometric primitive-type recognition and instance segmentation. The comparisons and ablations support its effectiveness under the reported validation setup, including the added color screening and latent relation constraint.

Critical weaknesses
  1. The core idea is to encode instance labels as colors and then predict the segmentation with a generative model. PartGen [1] and NaTex [2] are existing precedents. The technical route is also very close to SegviGen [3], including the reuse of TRELLIS.2, a frozen material codec, random-color instance encoding, and conditional flow prediction. The additions are mainly primitive-type encoding, color screening, and a local relation constraint. I see these changes as engineering adaptations for the CAD task, and not as a substantial methodological contribution.

    In Figure 2, the pipeline is packed too tightly and the font is too small. At a normal reading scale, the flow and the key modules are hard to make out. This is the main method figure, and it should not have this basic problem. Parts of the main text are also obscure and take repeated reading to follow.

  2. The paper emphasizes that the structured latent reduces dependence on how the input is discretized. The main evaluation uses only uniformly remeshed inputs, with no quantitative comparison across tessellations, mesh densities, or scan noise. The scan study shows only six qualitative examples, without metrics or baselines. The mechanism ablations use only a small training setup, which is not enough to show what the added designs contribute under full training.

[1] Minghao Chen et al. PartGen: Part-level 3D Generation and Reconstruction with Multi-view Diffusion Models. CVPR, 2025. Paper.

[2] Zeqiang Lai et al. NaTex: Seamless Texture Generation as Latent Color Diffusion. arXiv:2511.16317, 2025. Paper.

[3] Lin Li et al. SegviGen: Repurposing 3D Generative Model for Part Segmentation. arXiv:2603.16869, 2026. Paper.

Questions for authors
  1. The reported ablation uses only 9k samples and 10k steps, while the main experiment uses 273k samples, roughly a 30-fold difference in sample count. Primitive-type encoding, color screening, and the local relation constraint are the main additions relative to the existing route, so gains at this smaller scale do not show that they still help in the main experiment. The paper does not report the number of steps, hardware, or runtime of the full training, and I cannot tell whether an ablation closer to the main setting is infeasible. Please provide this information. If full training is indeed too expensive, please at least compare these designs against the version without them in a setting closer to the main experiment. Otherwise, it is hard to attribute the lead in the main table to these designs.

  2. The paper treats stability across discretizations as a main motivation, but the main table compares only uniformly remeshed inputs, and the scans are six qualitative examples. I therefore cannot treat “the result remains stable under a different mesh” as an established finding. Please report primitive-type recognition and instance-segmentation scores for the same CAD models under different tessellations, mesh densities, and scan noise, together with baseline comparisons. If these scores are clearly lower than those in the main table, this claim cannot count as a verified contribution.

Reviewer_agLxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses primitive type and instance segmentation of CAD surfaces by adapting pretrained TRELLIS.2 representations to predict per-face labels. It encodes primitive types and instance colors into a shared field and introduces explicit color separation, VAE-response gating, and latent relation supervision to reduce mixing between adjacent instances. Evaluation includes segmentation metrics, ablations, computational costs, and qualitative results. The method produces surface labels and instance partitions rather than fitted geometric parameters or complete CAD models.

Critical strengths

Primitive types and instance partitions are useful inputs to subsequent fitting: the type determines which surface model to fit, and the instance identifies which faces belong together. Producing both through one prediction pipeline is a clear and potentially useful goal for CAD reverse engineering. Strong gains under the shared validation protocol.

Critical weaknesses
  1. Final results are not evaluated on an independent test set. The same 1k validation set is used to select hyperparameters, checkpoints, and readout settings and to report final performance. Although baselines follow the same protocol, repeated selection may inflate the reported scores, limiting conclusions about generalization.

  2. Robustness to remeshing and scans remains insufficiently demonstrated. Quantitative comparisons primarily use uniformly remeshed inputs. The face-count experiment measures runtime rather than segmentation accuracy, while scan results lack ground-truth metrics and baseline comparisons. These experiments therefore do not establish superior robustness across discretizations or real-world scans.

  3. Missing close comparisons limit attribution of the gains. SegviGen, which shares the pretrained representation and palette-based generative segmentation approach, is not directly compared. A direct type-and-instance predictor using comparable pretrained features is also missing. Existing initialization and E/I/L ablations do not fully separate the proposed design's contribution from that of the pretrained representation, particularly for primitive recognition.

  4. Small-instance preservation and final grouping reliability are insufficiently evaluated. Without ground-truth palette reconstruction through the frozen VAE, representation loss cannot be separated from prediction error. Sensitivity to the final color threshold and inference randomness is also unclear. Direct measurements of boundary accuracy, small-instance recall, and merge/split errors are needed to establish reliable instance recovery.

  5. Benefits for CAD reverse engineering remain unverified. Higher segmentation scores do not necessarily imply more accurate primitive parameters or fewer fitting failures, especially when boundary errors persist. A shared downstream fitting evaluation would substantiate this motivation. Full CAD reconstruction is not required, but the claimed benefits should otherwise remain framed as potential applications.

Questions for authors
  1. Could the authors evaluate the fixed model on CAD designs excluded from model selection to better establish its generalization ability? It would also be helpful to clarify whether the scan data overlap with the training or validation sets and how instances are defined in the annotations.
  2. How robust is the method to variations in tessellation, mesh density, and noise? Could the authors provide comparisons using common surface samples or area-weighted metrics to avoid biases caused by mesh discretization, and clarify the input point counts and voxel resolution used in the experiments?
  3. Could the authors compare the method with SegviGen and a direct type-and-instance predictor under comparable settings? Further ablations would also help clarify whether the improvements come from generative prediction, joint encoding, or the proposed E/I/L components, rather than differences in pretrained representations, effective resolution, or training budget.
  4. How accurately does the frozen VAE preserve the ground-truth instance colors and boundaries, particularly for small instances? Could the authors analyze the sensitivity of merge/split errors to readout thresholds, resolution, and random seeds, and clarify how reserved codes, VAE-response calibration, and palette-sampling failures are handled?
  5. Since the method outputs surface labels and instance partitions rather than fitted geometric parameters or a complete CAD model, could the authors evaluate the resulting reconstruction quality using a consistent primitive-fitting pipeline across methods?
Reviewer_gUMHOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors propose a framework that segments CAD surfaces by geometric primitive type (plane/sphere/cylinder/etc.). They propose a novel method repurposing the RGB texture colors generated by TRELLIS 2 as a segmentation label. The original TRELLIS 2 has six channels allowing PBR material generation. The authors propose to repurpose these channels for segmentation instead, by reserving the first three RGB channels for instance segmentation and the last three for the CAD primitive label. Authors add specialized losses to explicitly separate RGB codes, and randomize colors so that the target is permutation invariant.

Critical strengths

The approach is simple and makes sense. Repurposing the texture layer of TRELLIS is an interesting approach for instance segmentation that definitely merits exploration. The figures are overall well-presented and clear.

Critical weaknesses
  • The authors claim that no AI has been used to write the paper (or even implementing the code). While I have no way of verifying claims regarding the code, I highly doubt that this paper has not been written using AI. Many, many passages clearly indicate typical AI writing, from them this particularly obvious passage: l.206: "This is a global codec-response proxy; it is not a local boundary-separability measurement."

I am not sure why the authors would conceal this information, considering that AI-written papers are within ICLR guidelines provided that usage is disclosed. It is also highly doubtful that anyone would not be using AI to write code in a 2026 ICLR submission, but again I have no way of verifying that claim.

  • This paper appears to implement a very general solution for 3D instance segmentation. None of the losses proposed in this paper are CAD-specific. Why was it evaluated solely on the CAD domain? Why not also explore classical part instance segmentation tasks and compare performance there with baselines. Known datasets can be utilized for this task and baselines already use them, e.g. PartObjaverse, PartNet or 3DCoMPaT++ to only cite a few. The advantage is also that retraining the baselines is not necessary anymore - as we can directly compare to the reported results in the original papers.
Questions for authors
  • Can you confirm that no AI has been used to write this paper or even write code? These claims are highly doubtful to me.
  • Can you confirm that the methods that you compare to (e.g. PartField, P3-SAM) are finetuned from the original weights rather than trained from scratch - using the author's recommended hyper-parameters? I am not sure what explains the huge performance gap between P3-SAM on this task and their reported performance on part instance segmentation. While CAD segmentation is a challenging task, this gap is really suspicious to me. Are the baselines trained until convergence under the same settings as the main paper?

Overall, I would be more confident in my assessment of this paper if authors provided results on PartObjaverse-Tiny or 3DCoMPaT++. Since none of the proposed losses are specific to the CAD domain, comparing this method to the original domain of the discussed baselines makes a lot more sense.

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

Related papers

Discussion (0)

Sign in to comment.