acceptodds
Withdrawn from ICLR 2027

DisCoMesh: Efficient and Flexible Artist-Mesh Generation via Sparse Computation powered Discrete Diffusion

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Meshes used in games and interactive 3D applications require not only accurate geometry, but also clean topology, fine details, and editable semantic components. However, existing generative methods struggle to produce such assets within a practical computational budget. To address this challenge, we introduce DisCoMesh, a framework based on the view that a multi-part artist mesh is better represented as a set of parts coupled by sparse global relationships rather than as a single global sequence or an ordered sequence of parts. Each semantic part is canonicalized and quantized independently using the full coordinate vocabulary, preserving local geometric detail across parts of different scales. A shared discrete diffusion model then denoises all parts simultaneously, removing the need for an artificial part-generation order. To coordinate simultaneous part generation, we introduce a global interaction module that integrates part placement and global geometric context with sparse token-level communication across parts, preserving coherent connections and spatial relationships. This part-wise formulation enables flexible control over parts, supporting selective part regeneration while preserving the rest of the mesh. Experiments demonstrate that DisCoMesh outperforms existing methods in generation quality, maintains competitive inference speed, and supports diverse downstream applications.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Generation, Artist Mesh Generation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ManeOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents DisCoMesh, a method for generating artist-created meshes that represents a point cloud shape as a set of semantic parts and generates all parts simultaneously through discrete diffusion denoising with an hourglass architecture. Unlike existing methods, it introduces a Global Interaction Module (GIM) to preserve dependencies between parts. GIM efficiently maintains inter-part consistency using bounding box conditioning, global point cloud conditioning, and sparse inter-part attention over compressed global summaries, similar parts, and adjacent parts.

Critical strengths
  1. The paper identifies limited modeling of inter-part relationships as a limitation of existing part-based generation methods and proposes a sparse attention mechanism to efficiently capture these relationships.

  2. The method generates highly detailed artistic meshes and demonstrates strong quantitative performance and fast generation speeds.

  3. The paper is clearly written and easy to follow, with informative figures and well-designed experiments. The impressive agent-based editing application further demonstrates its potential for broader use.

Critical weaknesses

Limited Originality

The proposed method largely reads as a well-executed combination of ideas from existing part generation research and artistic mesh generation methods. Specifically:

  • The discrete diffusion modeling framework resembles that of TSSR.
  • Part-aware self-attention, cross attention, global attention is largely used in part and scene generation work. (PartCrater, MIDI, etc.)
  • The hourglass architecture has appeared in Meshtron, FlashMesh, and iFlame.
  • The proposed Part-Aware Sparse Self-Attention appears to adapt Native Sparse Contextual Attention in MeshRipple by operating on parts rather than triangles.

Trade-off between efficiency and joint generation

The motivation of this paper is prior part generation methods inadequately model relationships between parts. However, limiting inter-part interactions is a deliberate design choice to reduce the token length and support larger meshes by generating part-by-part. Although PSSA reduces memory usage, the proposed method still appears to require generating all parts jointly, raising questions about its scalability compared with methods that generate parts independently.

Minor weakness

The paper should explicitly acknowledge that its Part-Aware Sparse Self-Attention is inspired by DeepSeek’s Native Sparse Attention and MeshRipple's Native Sparse Contextual Attention.

Questions for authors
  • Could the authors clarify how their approach goes beyond a direct integration of existing techniques?

  • What is the maximum number of faces the proposed method can generate, and how does this compare with other methods?

  • How does DiscoMesh address this limitation while retaining the computational benefits of part-based generation?

  • For symmetric parts, such as the left and right arms, does contextual attention capture their relationship when nearest-neighbor attention does not? Providing attention maps illustrating this would help clarify the benefits of contextual attention.

  • How does DiscoMesh handle the boundary parts during both inference and editing?

2 comments

Thank you for your review. We apologize for any ambiguity in the original manuscript. We hope our clarification of the method’s originality and our additional experimental results address your concerns.

W1 / Q1. Motivation for joint part generation. We would like to clarify the motivation and novelty of DisCoMesh. As discussed in Appendix G and illustrated in Fig. 20 of the revised manuscript, we observe substantial differences in local topology and geometry across semantic parts, with no clear ordering dependency among them, as also noted by Reviewer S6CP. These observations motivate our part-wise modeling and joint denoising formulation, in which all parts evolve simultaneously and PASSA coordinates their generation through sparse cross-part attention. DisCoMesh achieves the best generation quality among the evaluated methods while maintaining competitive inference speed (Table 1 and Fig. 13 of the revised manuscript).

Distinctions from prior work. PASSA changes both the unit of interaction and the rules for selecting which parts communicate. MeshRipple operates at the triangle level, whereas the content-adaptive route in PASSA uses part-level content to determine which parts interact. This allows a part to attend to another part even when the two are spatially distant. The spatial route introduces a 3D inductive bias by defining neighbors through the adjacency of part bounding boxes, in contrast to NSA’s sliding window over sequence positions. The Hourglass Transformer is a natural choice for our backbone because its hierarchical compression is well suited to modeling mesh sequences. Our contribution lies in how DisCoMesh generates parts simultaneously and coordinates them through content-adaptive and spatially grounded sparse attention. We have clarified these design differences and cited NSA and MeshRipple in Appendix F.4.

W2 / Q2 / Q3. Scalability of joint and multi-pass generation.

Scalability. DisCoMesh scales well when all parts are generated jointly. The largest mesh we generated in a single joint run contained 43,655 faces across 141 parts. Peak GPU memory usage was 43.66 GiB, leaving substantial memory headroom. For comparison, PartDiffuser evaluates on meshes with fewer than 4,000 faces, while MeshMosaic reports generating meshes with over 100K triangles through autoregressive patch generation.

Multi-pass generation. To scale beyond the size of a single joint run, a natural extension is to adopt a grouping strategy inspired by PartPacker [4]. The parts could be partitioned into groups and denoised over multiple passes, with all parts in each group generated jointly. This would limit the number of tokens processed in each pass while retaining parallel denoising within each group, providing a way to target meshes with over 100K faces.

Q4. Attention between symmetric parts. The attention map in Fig. 18 in the revised manuscript shows that the content-adaptive route captures interactions between symmetric parts even when they are not direct spatial neighbors.

Q5. Part boundaries during inference and editing. During training, the inter-part loss uses vertices shared by adjacent parts in the ground-truth mesh to encourage boundary alignment. At inference, corresponding boundary vertices tend to lie close together in the global frame and can be stitched using a rule-based merge.

During editing, we provide the preserved parts to the agent as spatial context and instruct it to generate voxels that connect to those parts at the intended interface. The 3D shape generator uses these voxels to produce a reference shape for the edited part. The agent then calls DisCoMesh to generate its topology from that reference shape. This guides the edited boundary into alignment with the preserved parts while keeping their geometry and topology unchanged, as illustrated in Figs. 5 and 6 of the revised manuscript.

[4] Tang et al. Efficient Part-level 3D Object Generation via Dual Volume Packing. 2025. https://arxiv.org/abs/2506.09980

Reviewer_ManeOpenReview ↗

Two main weakness remain insufficiently addressed.

Further Clarification Requested

Thank you for the response and additional results. However, my two main concerns remain insufficiently addressed.

  1. Originality. The response restates how PASSA differs from NSA and MeshRipple, through part-level units and bounding-box adjacency, but it does not explain why this is more than an adaptation of existing techniques. What specific technical challenge in joint part-wise generation do existing methods fail to address? Is there evidence that a straightforward integration is insufficient, such as a comparison with NSA’s sliding-window attention applied to parts or with dense cross-part attention?

  2. Scalability. Only peak memory is reported for the 43,655-face result. Can DisCoMesh maintain its generation quality and efficiency at the scale of MeshMosaic (100K+ faces), ideally compared on more complex shape? In addition, the multi-pass strategy remains unvalidated. Since joint generation is a core claim, how would parts in different groups interact and remain consistent with each other? If this cannot be validated, it should be clearly stated as a limitation.

Reviewer_S6CPOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DisCoMesh generates artist-style meshes by jointly and parallely denoising multiple parts of a whole object. Each part uses local coordinates and is conditioned on local point-cloud, while sparse inter-part attention and consistency losses help the parts fit together. The paper validates the method through baseline comparisons, ablations, and demonstrate generation up to 40K faces, and examples of part-level editing and face-budget control.

Critical strengths
  1. Efficiency and balanaced performance: Parallel part-wise discrete diffusion seems much more efficient than several AR baselines (Table 2), and it is nice to see 40K-face generation in Figure 10. Compared with PartDiffuser, dropping the inter-part generation order seems very reasonable: I do not see an obvious order in which semantic parts should be generated (lines 71–79). The numerical results are good, but the scale and part-level control are the more interesting parts for me.

  2. General design choices make sense to me: Intra-part attention and inter-part communication look reasonable to me. Local coordinates also make sense for small parts. I like that the losses take both intra-part connections and inter-part seams into account (lines 315–323). There are quite a few baselines, the visual results look strong, and the editing and face-budget examples are useful.

Critical weaknesses
  1. One thing I worry about is that the part information is already decided before sampling. PartSAM gives the parts and bounding boxes (lines 194–198), and each part has a fixed face budget (lines 324–325). This could be brittle when segmentation gets something wrong. The data also filter out imbalanced decompositions (lines 749–752), leaving less evidence for difficult inputs. Table 5 reports success rates, but quality is measured only on successful outputs (lines 1180–1183). A few failure examples would help me understand the practical limits. Unfortunately, I did not find not find the failure case / limitation analysis within the script.

  2. I am also not sure whether the parts are actually stitched together or just very close to each other. MSVD (lines 424–430) and the inter-part loss (lines 1088–1112) tell us about geometric agreement, but not necessarily vertex welding or face connectivity at intended joints. A seam can look fine but still need repair. This matters for mesh quality, and I would like it clarified.

  3. Comparison with Tripo P1 and Meshy T2 (Meshy SmartTopology) Since it is an efficient method of generating arctic-style mesh, comparing with Tripo P1 and Meshy SmartTopology will be very informative. One thing I am worried about is that these commercial models, based on diffusion and SpaceMesh kind of topology decoder, seem to have natural/native part segmentation ability.

Questions for authors

I would like to get some clarifications from the authors:

  1. How are per-part face budgets chosen in your experiments: reference face counts, a heuristic, or manually? How do you divide a user-specified total budget among parts (lines 324–325)? Do you have examples where merged/split parts, missed thin components, or uneven budgets cause generation to fail?

  2. When computing MSVD, how do you find the seam vertices in the generated parts? Is there any welding or repair after sampling? I would like to know whether intended joints are connected before repair, not just close in space (lines 424–430).

  3. Figure 10 is very informative. What sampling steps, remasking settings, and peak GPU memory were used for the 40K-face runs? Timing excludes input preparation and repair (lines 1185–1187), so how much time does PartSAM add? The appendix says approximately 1B parameters (lines 862–863), but I could not find the hidden width, depth, attention heads, or GIM placement. Please also give the noise schedule, inter-part loss weight, and how the 4K multi-part training steps are split between the two GIM phases (lines 993–996).

  4. One thing I am concerned about is the eternal topic about "DLLM vs strict causal LLM". Based on my own experience working on Block Diffusion, I always saw a strict performance degradation (of course here we are not talking about lossless speculative decoding). May I know your opinion on this? I wonder whether some quality is traded for parallel decoding in this project.

Suggestions for discussion or future work:

The MSVD discussion also made me think of related local and part-level editing work, including Masked LRMs, PartGen, and MeshFIM. A short discussion of how these methods handle boundaries and preserve the unedited region would be nice.

3 comments

We appreciate the reviewer’s recognition of our motivation for generating parts without a fixed order and their positive assessment of our design choices. In the revised manuscript, Fig. 20 illustrates how local topology and geometry vary across parts of the same mesh, and Appendix G explains how this observation motivates our joint part-wise generation framework.

W1 / Q1 — Face budgets, segmentation robustness, and failure cases

Robustness to face budgets. DisCoMesh can generate high-quality meshes without requiring expert knowledge to assign an appropriate face budget to each part. For the Artist results in Table 1, we use the reference face count of each part. For Dense, we manually set a total budget of 10K faces and distribute them according to each part’s bounding-box surface area. At a fixed total face count, reconstruction quality remains comparable to the reference allocation when faces are allocated by bounding-box surface area or when the reference part budgets are randomly perturbed (see the table below). Fig. 15 of the revised manuscript (Fig. 11 of the original manuscript) illustrates how the level of detail changes for the same shape across four face budgets.

Face allocation across partsCD-L1 ↓EF1 ↑
Reference face counts0.00820.676
Heuristic budgets (bbox surface area)0.00860.636
Randomly perturbed reference budgets (0.7–1.3×)0.00830.670

Segmentation errors and data filtering. DisCoMesh can still generate a coherent mesh when the input segmentation has noisy boundaries between parts, as shown in Fig. 19 of the revised manuscript. During dataset construction, we use imbalance statistics to flag potentially erroneous decompositions. As illustrated in Fig. 9 of the revised manuscript, our filtering removes incorrect segmentations and low-quality data while retaining small, intricate, and meaningful parts.

Full-test-set evaluation and limitations. DisCoMesh achieves the best scores on geometric metrics on both Artist and Dense when the HD filter is removed (see the tables below). Table 5 of the original submission reports success rates over the full test set and averages geometric metrics over successful outputs (valid meshes with HD ≤ 0.15) to prevent extreme errors from dominating the averages.

Artist Meshes

MethodCD-L1 ↓HD ↓NC ↑EF1 ↑
DeepMesh0.02820.11820.8440.317
FastMesh0.02430.08830.7860.121
Mesh Silksong0.01900.10280.8390.216
MeshMosaic0.01650.17510.8650.392
MeshRipple0.02330.13630.8320.228
MeshFlow0.01610.04480.7970.082
TSSR0.01000.06850.9130.517
LATO.20.01170.05070.8340.196
DisCoMesh0.00780.03110.9340.681

Dense Meshes

MethodCD-L1 ↓HD ↓NC ↑EF1 ↑
DeepMesh0.11330.39770.7370.045
FastMesh0.06110.27210.6850.015
Mesh Silksong0.07250.30650.7290.029
MeshMosaic0.02060.13330.8290.132
MeshRipple0.04690.24570.7100.020
MeshFlow0.01950.05510.7540.012
TSSR0.01260.06570.8940.108
LATO.20.01300.05230.8230.068
DisCoMesh0.01010.04920.9210.263

Failure cases are rare. When many thin, disconnected sheets are tightly interleaved, DisCoMesh may struggle to reconstruct the individual sheets.

W2 / Q2. Part boundaries and MSVD calculation. Each part is generated in its own canonical coordinate frame, so corresponding boundary vertices may have slightly different coordinates when mapped back to the global frame. During training, the inter-part loss uses vertices shared by adjacent parts in the ground-truth mesh to encourage alignment across part boundaries. As a result, corresponding boundary vertices tend to lie close together after sampling, so the boundaries can be stitched together by merging vertices within a distance threshold. We evaluate the meshes without welding or repair, so the raw output does not guarantee shared vertices at intended joints.

To compute MSVD, we use the ground-truth seam locations to identify nearby open-boundary vertices in each generated part. We then match each identified vertex to the nearest seam vertex on the neighboring generated part and average the distances between matched vertices.

W3. Commercial comparisons and part structure. We could not find publicly available model weights or inference code for Meshy T2 or Tripo P1, making a fair comparison difficult. Nevertheless, as suggested, we conducted a qualitative comparison with both commercial systems through their official APIs. On the examples shown in Figs. 11 and 12 of the revised manuscript, DisCoMesh achieves comparable or even better results in shape completeness and geometric detail.

Meshy T2 natively supports multi-part meshes whose connected components emerge from the generated topology [1]. Tripo P1 does not support the “generate_parts” option in its generation API [2]. Tripo provides a separate semantic segmentation endpoint for existing meshes [3]. DisCoMesh uses semantic parts identified by PartSAM on the reference geometry as its units of generation and editing. Users can therefore regenerate a selected part while preserving the exact geometry and topology of every other part.

[1] Meshy T2 technical report — https://arxiv.org/abs/2607.28675

[2] Tripo official API documentation — P1 parameter support — https://developers.tripo3d.ai/en/docs/changelog

[3] Tripo Mesh Segmentation API — https://developers.tripo3d.ai/en/docs/mesh-segment

Q3. Inference and training details. For the 40K-face runs in Fig. 13 of the revised manuscript (Fig. 10 of the original manuscript), we used 50 hybrid sampling steps, with the classifier selecting which tokens to remask at each step. Peak GPU memory usage was 43.66 GiB on a mesh with 43,655 faces and 141 parts. PartSAM preprocessing took 37 seconds on average for input meshes with approximately 40K faces.

DisCoMesh has approximately 1.14B parameters. Its Hourglass Transformer has a hidden width of 1,536 and 20 blocks across five stages with depths [2, 4, 8, 4, 2]. Each block uses 12 attention heads. We insert GIM after PACA in selected blocks within each hourglass stage. During training, we sample uniformly from 1,000 discrete timesteps. At normalized time t, each token is retained with probability t² and otherwise replaced by a mask token. The inter-part Wasserstein loss has a weight of 10.

Q4. Does parallel decoding trade off mesh quality? We appreciate the opportunity to discuss this question. As shown in Table 1 and Figs. 4 and 10 of the revised manuscript, DisCoMesh generates all parts simultaneously and achieves the best reconstruction quality among the evaluated methods. Thus, we hold that parallel decoding does not inherently require sacrificing mesh quality. In language, word order helps convey meaning, giving causal models a natural sequence to follow. In artist meshes, we observe no clear dependency that requires one part to be completed before another can be generated (Appendix G, Fig. 20 in the revised manuscript). We therefore denoise the parts jointly, allowing them to evolve together and exchange information through PASSA.

To measure the additional inference time required by sequential generation, we implemented a sequential variant that uses the same backbone to denoise one part at a time. At 8K faces and 50 denoising steps, this variant takes 255.70 s on average even with KV caching, compared with 109 s for DisCoMesh.

Suggestion. Local and part-level editing. Thank you for pointing us to these works. In Appendix F.3 of the revised manuscript, we now discuss how Masked LRMs, PartGen, and MeshFIM handle edit boundaries and preserve unedited regions. We also clarify that DisCoMesh preserves the exact geometry and topology of every unedited part, while the inter-part loss encourages alignment along the regenerated boundary.

I sincerely thank the authors for addressing my concerns directly and honestly in their rebuttal response above.

Face budget assignment

The authors have made it clear how face budget assignment works in their method. However, using either the reference face counts from the corresponding GT low-poly mesh or a fixed total face budget distributed according to AABB surface area ratios seems somewhat heuristic and not very general.

Segmentation errors

Thanks for the additional information on data filtering. I think I have enough information regarding this part. Given the use of post-processing to handle boundary vertices in this paper, I do not think this should be a major issue here.

Part boundaries and MSVD calculation

I appreciate the authors' honesty in pointing out that they handle the boundary issue through post-processing, and I believe this is a fair approach.

Commercial comparisons

Thanks for your efforts in comparing with Tripo P1 and Meshy T2. For this part, I do have some further questions.

(1) Firstly, for a fair comparison with Tripo P1 and Meshy T2, I would like the authors to consider the following receipe: For DisCoMesh, do not use GT point clouds from low-poly meshes. Try using the PC from a high-poly generation method (such as Sparc3D, which Hi3D has just released its open-source project on GitHub or you can also directly call their APIs) with the same input image used for Tripo and Meshy. In this way, we can eliminate the bias from using GT low-poly point clouds and gain a better understanding of the comparison.

(2) Tripo P1 does not support ... Although there is no clear API, part segmentation information can be retrieved from connectivity information. The argument that DisCoMesh uses sema... is not clear to me. If Tripo P1 and Meshy T2 can generate parts, and their training data are properly preprocessed, the quality of part generation should not be an issue.

Trade-off between parallel decoding and mesh quality I appreciate the information here and do not have further question for now.

Overall, thanks to the authors' rebuttal, I now have a better picture of this project. I do expect the authors to further address my concerns about the comparison settings with commercial methods. Nevertheless, I still think this work has its merits and would like to champion its acceptance at ICLR.

Reviewer_pBwqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DisCoMesh generates artist meshes conditioned on a reference shape by denoising all semantic parts simultaneously. Each part is normalized and quantized in its own coordinate frame and represented by face-wise coordinate tokens. A shared Hourglass Transformer combines local self-attention and point-cloud conditioning with a Global Interaction Module using bounding boxes, global geometry, and sparse cross-part attention. An additional interface consistency loss encourages adjacent parts to align. The paper reports improved geometric fidelity and sharp-edge recovery over the evaluated baselines, faster inference than several high-quality baselines, and applications to selective part regeneration and local mesh-complexity control.

Critical strengths

1.DisCoMesh achieves better reconstruction quality than existing methods while being considerably faster, demonstrating its effectiveness for explicit discrete mesh generation. 2. The proposed representation enables flexible local editing while preserving the geometry and topology of unmodified parts. The ablation study also shows that the proposed interface coordination helps improve consistency between parts.

Critical weaknesses
  1. The methodological advance over the closest prior work is limited and insufficiently isolated. MeshMosaic already establishes decomposition, independent local quantization, and local/global geometric conditioning; PartDiffuser already combines part-wise generation with discrete diffusion within parts. The main change is to denoise parts jointly, with sparse communication replacing sequential conditioning. This is a reasonable extension, but the paper does not establish how much its gains specifically require this formulation. PartDiffuser is cited but absent from quantitative comparisons, and there is no controlled comparison between sequential and simultaneous part diffusion with matched data, backbone, and sampling budget. The sparse module is ablated only as a whole; the necessity of its three interaction routes and gated fusion is unclear. Consequently, the evidence does not yet establish a sufficiently strong contribution beyond combining and extending these existing ideas.
  2. Missing recent comparisons weaken the claims about efficiency and structured artist-mesh generation. Nexus [1], Meshy T2 [2], and Strips as Tokens (SATO) [3] should all be cited and discussed. Quantitative and qualitative comparisons with Meshy T2 and SATO are particularly important. Meshy T2 directly produces multi-component assets, while SATO jointly generates meshes and UV-chart partitions; these outputs differ from prescribed semantic parts, but both are relevant to claims about structured and editable meshes. Meanwhile, 109 seconds for an 8K-face mesh is still substantial. Meshy's official optimized-system report [4] gives approximately 1 second for generation and 2 seconds end to end; its paper reports 3 seconds for retopology and 6 seconds for image-to-mesh generation. These are different protocols, not a matched speedup measurement, but they make a controlled comparison essential. Even Table 2 already has MeshFlow at 3.68 seconds. Without the missing baselines, explicit sampling settings, and quality-time curves, the practical significance of the efficiency contribution remains unclear. [1] Wang et al. Nexus: Native Mesh Generation with Diffusion. 2026. https://arxiv.org/abs/2607.13563 [2] Xu et al. Meshy T2: Fast Native Mesh Generation with Flow Matching. 2026. https://arxiv.org/abs/2607.28675 [3] Xu et al. Strips as Tokens: Artist Mesh Generation with Native UV Segmentation. 2026. https://arxiv.org/abs/2604.09132 [4] Meshy. Meshy T2: Native 3D Mesh Generation in 2 Seconds. August 12, 2026. https://www.meshy.ai/blog/meshy-t2-native-3d-mesh-generation
Questions for authors
  1. What measurable benefit comes specifically from simultaneous rather than sequential part diffusion? Can you provide a matched comparison with PartDiffuser or a sequential version of your own model?
  2. Can you compare against Meshy T2 and SATO at comparable output budgets, reporting geometry, sharp-edge recovery, topology, structured outputs, and runtime? Please distinguish connected components and UV partitions from semantic parts, and state any differences in conditioning inputs.
  3. How many denoising iterations and backbone evaluations produce the 109-second result? Please report how runtime is distributed across local processing and global interaction, and show the quality-time trade-off as the sampling budget varies.
  4. Are all three sparse interaction routes and their learned fusion necessary? Comparisons with neighbor-only interaction and simpler fusion, including time and memory, would directly affect my assessment of the architectural contribution.
1 comment

Thank you for recognizing our reconstruction quality and part-level editing capabilities. We address your concerns below.

W1.1. Motivation and novelty. We argue that DisCoMesh is more than a simple extension of existing methods. As discussed in Appendix G and illustrated in Fig. 20 of the revised manuscript, we observe substantial differences in local topology and geometry across semantic parts, with no clear ordering dependency among them. This observation motivates us to model each part in its own coordinate frame and denoise all parts simultaneously, using part-aware sparse self-attention (PASSA) to exchange information across parts during denoising. As shown in Tables 1 and 2 of the revised manuscript, DisCoMesh achieves the best reconstruction quality among the evaluated methods while maintaining competitive inference speed. Beyond reconstruction, this formulation allows users to regenerate a selected part while preserving the exact geometry and topology of every unmodified part.

W1.2 / Q1: simultaneous vs sequential part denoising

We could not find publicly available model weights or runnable inference code for PartDiffuser. We therefore implemented the sequential variant suggested by the reviewer, using the same backbone, discrete mesh representation, and model size as DisCoMesh. Under the same 8K-face setting with 50 denoising steps, the sequential variant took 1,630.04 s on average without KV caching and 255.70 s with KV caching. DisCoMesh took 109 s on average. Even with KV caching, sequential generation required more than twice as long. Simultaneous denoising reduced generation time by approximately 57.4% at the same sampling budget. Through simultaneous denoising and sparse cross-part interaction, DisCoMesh also demonstrates comparable or even better scalability on meshes with high face and part counts. On the same sample with 43,655 faces and 141 parts, DisCoMesh used 43.66 GiB of peak GPU memory, while the sequential variant with KV caching used approximately 45.8 GiB.

W1.3 / Q4. Sparse attention routes and learned fusion All three sparse attention routes and their learned fusion contribute to the final mesh quality. As shown below, removing any route or replacing learned fusion with equal-weight averaging degrades geometric quality.

VariantCD-L1 ↓CD-L2 × 10³ ↓HD ↓NC ↑EF1 ↑
Full model0.00750.0370.0240.9130.693
w/o compressed global route0.00770.0400.0270.9130.690
w/o content-adaptive route0.00760.0390.0270.9120.688
w/o spatial neighbor route0.00760.0380.0260.9110.687
Equal-weight fusion0.00760.0380.0260.9110.684

W2.1 / Q2. Recent baselines and part structure We could not find publicly available model weights for Nexus, Meshy T2, or SATO, making a fair comparison difficult. Nevertheless, as suggested, we conducted a qualitative comparison with two commercial systems, Meshy T2 and Tripo P1, through their official APIs. DisCoMesh achieves comparable or even better results in shape completeness and geometric detail as illustrated in Figs. 11 and 12 of the revised manuscript. We have also added citations and discussion of Nexus, Meshy T2, and SATO in Appendix F.2.

Beyond the commercial comparisons, we also evaluated LATO.2 to broaden our comparison with continuous latent flow methods. DisCoMesh achieves the best results across all five metrics on both the Artist and Dense datasets. The EF1 results further support our observations in Appendix F.2.

DatasetMethodCD-L1 ↓CD-L2 × 10³ ↓HD ↓NC ↑EF1 ↑
ArtistLATO.20.01130.1110.04480.8350.200
ArtistDisCoMesh0.00770.0480.02810.9340.683
DenseLATO.20.01260.1130.04720.8240.068
DenseDisCoMesh0.01000.0790.04460.9200.267

Part structure. Meshy T2 can produce meshes with multiple connected components, while SATO generates meshes with UV-chart partitions. DisCoMesh uses semantic parts identified by PartSAM, allowing users to regenerate a selected part while preserving the exact geometry and topology of all other parts.

W2.2 / Q3. Sampling cost and quality–time trade-off With our default setting of 50 hybrid sampling steps, DisCoMesh generates an 8K-face mesh in 109 seconds on average. Local processing and global interaction take 80.37% and 19.63% of the time. Revised Fig.14 shows quality across sampling budgets. Even at 30 steps, DisCoMesh achieves SOTA-level quality in 65 s. Increasing from 30 to 50 steps improves quality, with only modest gains from 50 to 200.

While MeshFlow is faster, DisCoMesh substantially improves surface fidelity and sharp-edge recovery. Revised Fig. 13 (original Fig. 10) illustrates the quality–time trade-off across face counts.

Comments on the paper

1 comment

Dear Program Chairs, Area Chair, and Reviewers,

Thank you for your time and feedback. We summarize our clarifications and additional experimental results below.

1. Motivation and originality. DisCoMesh models parts locally to capture their distinct topology and geometry (Fig. 20 of the revised manuscript). We observe no clear ordering dependency among semantic parts. We therefore denoise all parts simultaneously, with PASSA coordinating their generation through sparse cross-part attention (Appendix G). In Appendix F.4, we clarify how PASSA uses part-level content and spatial neighborhoods to select interactions and how these choices differ from MeshRipple and NSA. (See pBwq W1.1 and Mane W1/Q1.)

2. Additional experimental results

2.1 Scalability. DisCoMesh can generate a mesh with 43,655 faces across 141 parts in a single joint run, with peak GPU memory usage of 43.66 GiB. At 8K faces and 50 steps, simultaneous denoising reduced average generation time by 57.4%, from 255.70 s for the cached sequential variant to 109 s for DisCoMesh. (See pBwq W1.2/Q1 and Mane W2/Q2/Q3.)

2.2 Effectiveness of individual PASSA components. All three sparse attention routes and their learned fusion contribute to mesh quality. Removing any one route or replacing learned fusion with equal-weight averaging degrades geometric quality. The attention map in Fig. 18 of the revised manuscript shows that the content-adaptive route captures interactions between spatially separated symmetric parts. (See pBwq W1.3/Q4 and Mane Q4.)

2.3 Comparison with additional baselines and commercial systems. With LATO.2 included, DisCoMesh achieves the best results on all five metrics on both Artist and Dense. In qualitative comparisons with Meshy T2 and Tripo P1, DisCoMesh achieves comparable or even better shape completeness and geometric detail on the examples shown. (See pBwq W2.1/Q2 and S6CP W3.)

2.4 Robustness to face budgets. DisCoMesh generates the same shape at different levels of detail across a range of face counts. At a fixed total face count, reconstruction quality remains comparable across the tested allocation strategies, including allocation by bounding-box surface area and random budget perturbations. (See S6CP W1/Q1.)

2.5 Quality–time trade-off. We report curves showing how generation quality varies with the sampling budget (Fig. 14 of the revised manuscript). Reducing the default budget from 50 to 30 steps cuts average generation time for an 8K-face mesh from 109 s to 65 s while maintaining reconstruction quality comparable to the state of the art. Increasing the budget from 50 to 200 steps yields only modest quality gains. (See pBwq W2.2/Q3.)

3. Implementation and editing details. We clarify the model architecture and sampling settings and report the time spent on local processing, global interaction, and PartSAM preprocessing (see pBwq W2.2/Q3 and S6CP Q3). We explain how preserved parts guide generation at the edit boundary (see Mane Q5) and discuss how DisCoMesh and prior editing methods preserve unedited regions (Appendix F.3).

We thank the reviewers and chairs for their time and constructive feedback. We have revised the manuscript to address the points raised and highlighted the updated text in blue. We welcome further questions and discussion and are happy to provide additional clarification.

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

Related papers

Discussion (0)

Sign in to comment.