acceptodds
Under review as a conference paper at ICLR 2027

Look Both Ways: Future-Informed Autoregressive High-Poly Mesh Generation

Abstract

Existing part-wise autoregressive 3D mesh generation methods predict each next part conditioned only on preceding parts. Without explicit future-informed topological compatibility guidance, early local topological biases can accumulate along the generation chain and eventually lead to topological drift in high-face-count settings. We propose FAMesh, a topologically consistent framework for high-face-count 3D mesh generation. FAMesh formulates target mesh generation as part-wise autoregressive prediction of topological sequences and introduces a global topological scaffold (TopoScaf), which captures the coarse global structure of the target mesh as a compact part-level topological sequence. TopoScaf complements the target-part autoregression with explicit cues about future-part topological compatibility, steering generation along a topologically consistent trajectory. Furthermore, we design Part-Structured Asymmetric Interaction (PSAI), which reformulates information transfer required for TopoScaf modeling and TopoScaf-to-target topological guidance into efficient interactions over discriminative part-level structural summaries. This design achieves a training speedup, without compromising TopoScaf modeling fidelity or its topological guidance for target generation. Experiments show that FAMesh outperforms existing methods in generation quality and topological consistency.

Primary area:
generative models
Keywords:
3D mesh generation, diffusion, autoregressive 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_6aYGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FAMesh, an autoregressive framework for high-poly mesh generation that aims to reduce error accumulation and cross-part structural inconsistencies in conventional part-based generation.

Critical strengths

1.The paper is clearly written, with well-explained equations and notation. The method overview is also intuitive and easy to follow.

2.This paper addresses an important problem in high-poly autoregressive mesh generation: cross-part structural inconsistencies caused by the accumulation of errors from previously generated parts. I think this problem is meaningful, and the proposed approach is both well-motivated and reasonably novel.

Critical weaknesses
  1. Minor Issues with Presentation

Some symbols and text labels in Fig. 3 overlap, making the figure somewhat difficult to follow. I suggest revising the layout to improve readability. Since the proposed method targets high-poly mesh generation, the generated meshes contain a large number of faces, making it difficult to inspect their topology even when zooming in on the figures, particularly in Fig. 5. The teaser includes several close-up views of individual parts, which are helpful. However, the faces in some regions are still too dense to clearly distinguish their arrangement. The wireframe rendering also resembles a typical MeshLab visualization, which is not a clear way for rendering wireframes. In addition, the rendering quality and resolution appear to be inconsistent across different parts(e.g. the zoom-in images of the teaser), with some close-ups looking noticeably clearer than others. At present, the teaser is probably the only figure where I can reasonably assess the model's topology generation quality and the generated topology seems less regular than that of some previous autoregressive artist-mesh generation methods. I understand that this paper primarily focuses on scalable high-poly mesh generation rather than retopology. Therefore, I do not consider this a major weakness of the work. Nevertheless, clearer visualizations would make it easier to assess the actual quality of the generated meshes.

  1. See the Questions part.
Questions for authors
  1. The data curation process establishes a fixed part ordering during training. However, it is unclear how Scafor identifies part boundaries within the flattened TopoScaf sequence during inference. Could the authors clarify how the model determines where one part ends and the next begins during inference?

  2. Intuitively, PSAI introduces a trade-off between training&inference efficiency and generation quality/information preservation by compressing token features into structural summaries through gated aggregation. However, Table 3 shows that PSAI not only substantially reduces training time but also improves several geometric quality metrics. In TopoDec, PSAI mainly compresses the scaffold features of previous parts, while in ScafPAR, it compresses global TopoScaf features to reduce cross-attention costs. Both operations reduce access to fine-grained topological information and will cause accumulated error. I would therefore appreciate a more detailed explanation of why such compression can improve generation quality rather than merely preserve it. Is there any theoretical justification or empirical analysis that could help explain this behavior?

  3. Why did the authors choose EdgeRunner as the mesh tokenizer? Based on the results in Fig. 5 and my experience with the open-source EdgeRunner code, it is not a good method for information preservation which suffers from many artifacts(e.g. missing faces), but I didn't see these issues in the results presented in this paper. Moreover, more recent tokenization methods(e.g.BPT and DeepMesh) have introduced block-wise and patch-wise compression strategies that achieve more compact mesh representations and their results are more stable after broad test in the 3D generation community. I hope authors discuss why these alternatives were not considered. If they can be adapted to the proposed framework, a comparison or ablation study using different tokenizers would help clarify whether EdgeRunner is particularly suitable for the proposed traversal-consistent generation framework.

  4. In practical part-based mesh generation, it is often desirable to keep the contact regions between adjacent parts open, rather than generating unnecessary closing faces for each part independently. I could not find a discussion of this issue in the paper. So I hope author s can discuss more about it.

  5. The paper reports using approximately 85K mesh assets for training. In my experience, this is a relatively modest dataset size for part-based mesh generation, particularly considering the quality of assets in Objaverse and the wide range of face number. However, the reported results are surprisingly not bad. I was particularly impressed by the last row of Fig.5, where the model appears to recover some Chinese character structures. Although the characters are not reproduced perfectly, the result is still interesting given the training data scale. I am wondering if the author do some extra practice e.g. data cleaning or data augmentation. I am truly interested in how the authors achieved this quality and apparent generalization with a relatively small number of mesh assets.

  6. Do the authors plan to release the training code and the processed dataset?

4 comments

We sincerely thank the reviewer for the detailed and constructive feedback. We appreciate your recognition of the paper’s clarity and of the importance, motivation, and novelty of addressing error accumulation and cross-part structural inconsistencies in high-poly mesh generation.

We believe our response has addressed most of your concerns and we respectfully ask if you're comfortable to re-consider the rating in light of our efforts and dedication. Below, we provide detailed responses to all the comments:


Q1: Overlapping labels and unclear, inconsistent wireframe renderings hinder readability and assessment of topology quality.

Thank you for the feedback. We revised Fig. 3 to resolve overlapping labels and re-rendered all visualizations using a unified pipeline to improve wireframe clarity and image consistency.

We identified meshes with poor topology before receiving the reviews. While awaiting the reviews, we refined the data filtering, expanded the training set to 120K meshes, and completed retraining. We updated Figs. 1, 5 and Tables 1–2 accordingly, with additional examples in Appendix Fig. 9.


Q2: How Scafor identifies part boundaries within the flattened TopoScaf sequence during inference

During training, we append EOS to each TopoScaf part subsequence and an additional EOS to the full sequence. At inference, TopoDec predicts these markers autoregressively: EOS ends the current part, a subsequent traversal-start token (ID 5) begins the next part, while a second consecutive EOS terminates the sequence.


Q3: Why does PSAI’s compression improve generation quality, and what theoretical or empirical evidence supports this?

We attribute the gains in Table 3 to task-aligned interaction granularity: fine-grained modeling within parts and structural summaries across parts. This lowers interaction costs and may reduce interference from redundant details and local prediction errors.

For TopoDec, parts are watertight connected components with independent vertex indices, assembled directly after generation. Fine-grained topology modeling therefore primarily serves connectivity within parts, while cross-part conditioning only needs to coordinate relative placement and geometric contact without requiring shared vertices. MeshMosaic [1] likewise weakening the influence of cross-part information, providing a rationale for concentrating fine-grained modeling only within parts.

For ScafPAR, each face comprises multiple tokens. Aggregating local tokens in topology traversal order can thus be viewed as forming face- or patch-level representations that providing explicit cues about position, surface shape, and orientation. Moreover, Hscafn\mathrm{H}_{\mathrm{scaf}}^n compliments additional topology information for the current part. FACE [2] offers related theoretical support: it encodes each face’s coordinates into a single token, shortening sequences while achieving high-quality reconstruction.

Aggregation versus scaffold resolution. Aggregation learns representations from an existing scaffold, reducing its resolution may remove essential geometric cues. The degradation in Table 7 underscores the need for sufficient scaffold resolution.

Notably, although one token error can damage multiple associated faces, the mild local scaffold corruption in Fig. 10 (second row) does not reappear after refinement. This observation is consistent with aggregation improving tolerance to local token errors.

Q4: Why choose EdgeRunner over BPT or DeepMesh, and how does tokenizer choice affect the framework’s performance?

Compatibility with coarse–fine traversal alignment:

The QEM collapse history assigns fine faces to coarse-face regions, which are processed in coarse traversal order. EdgeRunner’s L/R adjacency transitions and BOM restarts let us enforce this order without changing its vocabulary or decoding rules.

BPT and DeepMesh form patches around central vertices, whereas SATO [3] constructs strips. These groups may cross coarse-face boundaries. Adapting them is possible in principle but would require region constraints and potentially splitting patches or strips. EdgeRunner thus provides a more direct implementation of our alignment mechanism.

Different tokenizers use distinct serialization grammars, so direct substitution would cause a training–inference mismatch. These ablations therefore require retraining the model with each tokenizer. We are willing to include these experiments in the camera-ready version.

Tokenizer comparisons:

To assess information preservation, we compare four tokenizers, with results shown below.

R1-Table 1. Tokenizer comparisons. All methods use the same GT inputs.

MethodMean token lengthMean decoded facesHD ↓CDL1CD_{L1} ↓F1 ↑ECD ↓EF1 ↑
DeepMesh88,708.47033,399.4700.0130.0050.6590.0010.636
BPT32,104.17015,931.7300.0170.0070.4720.0020.468
SATO196,308.80058,836.2700.0110.0060.6620.0010.637
EdgeRunner220,389.73052,219.4300.0140.0040.6610.0010.633

EdgeRunner achieves geometric fidelity comparable to DeepMesh and SATO and better than BPT in this evaluation. The newly added Fig. 11 in Appendix further shows that EdgeRunner, DeepMesh, and SATO preserve overall shape but exhibit local degradation in fine structures, such as facial details. This degradation is not unique to EdgeRunner and is consistent with a shared limitation of coordinate quantization: nearby vertices can map to the same position, collapsing short edges and creating degenerate triangles. SATO’s official implementation also acknowledges this limitation. Together, these results and its compatibility with our alignment mechanism support our choice of EdgeRunner.


Q5: Why are contact regions between adjacent parts closed rather than left open?

We thank the reviewer for raising this point. Whether contact regions should remain open depends on the part definition and downstream use. In line with the independent part modeling approach of BANG[4], FAMesh represents each part sequence (S(n))(\mathcal{S}^{(n)}) as an independent watertight mesh, intentionally retaining closed contact surfaces. This allows topological closure within each part during autoregressive generation, without requiring boundary stitching between parts. Geometric coordination between parts is guided by the contours, layout, and contact relationships provided by TopoScaf. Meanwhile, complete surfaces also support independent use, repositioning, and reassembly, reducing the need to close and repair exposed boundaries when assembly relationships change. We have added this discussion to Section 3.1.

Q6: How does FAMesh achieve good quality and generalization with 85K assets, and were data cleaning or augmentation used?

We thank the reviewer for highlighting the missing training details.

Pretraining and data preparation:

Both TopoDec and ScafPAR are initialized from the EdgeRunner checkpoint, retaining its base architecture and serialization while incorporating our proposed modules. We use Objaverse-XL assets filtered by TRELLIS [5,6], with random rotations, point-cloud noise, image degradation, and topology-preserving scaffold perturbations.

The fine-mesh training set was subsequently expanded to 120K assets. TopoDec uses 200K assets simplified to approximately 4.5K-face scaffolds. The shorter sequences reduce training costs, allowing a larger dataset to broaden structural coverage and improve scaffold stability. For context, the base models of TreeMeshGPT [7], and FACE[2] also use training datasets of comparable scale. We have added the relevant details to Section 4.1.

Scaffold guidance reduces the burden of topology inference:

TopoDec establishes a compact global structure, and ScafPAR refines individual parts, limiting each prediction to the current part’s sequence. Aligning coarse and fine traversals provides explicit topological trajectories and cross-part structural cues, reducing the need to infer topology from surface geometry alone. The marked degradation without traversal consistency (Table 4) supports this design. This also helps explain the “武” example in Fig. 5: the scaffold provides the main stroke layout and structural cues, helping ScafPAR reconstruct geometric details more reliably.


Q7: Do the authors plan to release the training code and the processed dataset?

We thank the reviewer for their interest in our work. It would be our pleasure to make our training code, processed dataset, and preprocessing scripts publicly available upon acceptance to support further research.

[1] Scaling Artist Mesh Generation via Local-to-Global Assembly.

[2] FACE: A Face-based Autoregressive Representation for High-Fidelity and Efficient Mesh Generation

[3] Strips as Tokens: Artist Mesh Generation with Native UV Segmentation

[4] BANG: Dividing 3D Assets via Generative Exploded Dynamics

[5] Structured 3D Latents for Scalable and Versatile 3D Generation

[6] Native and Compact Structured Latents for 3D Generation

[7] TreeMeshGPT: Artistic Mesh Generation with Autoregressive Tree Sequencing

Dear Reviewer,

I hope this message finds you well. As the discussion period is nearing its end, I wanted to ensure we have addressed all your concerns satisfactorily. If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and we’re eager to address any remaining issues to improve our work.

Thank you for your time and effort in reviewing our paper!

Reviewer_qMDHOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

Authors propose FAMesh, a topologically consistent framework for high-poly 3D mesh generation. To handle meshes with large face counts, FAMesh first generates a coarse mesh scaffold for each part. It then independently generates a detailed mesh for each part, conditioned on its corresponding scaffold, and assembles the resulting meshes. This approach enables the generation of longer face sequences and more complex structures.

Critical strengths
  1. Previous approaches require tokens representing the entire shape or previously generated parts, making long sequences difficult to handle. The proposed method addresses this limitation by generating parts independently, enabling meshes with longer face sequences.
  2. To the best of my knowledge, this is the first work to use a simplified mesh, termed a scaffold, as conditioning input for part-level mesh generation.
  3. Design PSAI structure to guide part-level structural summaries that improves training and inference efficiency with comparable geometric performance.
Critical weaknesses

Lack of structural novelty

  • The proposed method appears similar to MeshMosaic, which also uses part-level points to guide generation. The main distinction appears to be the use of scaffolds instead of point clouds as conditioning inputs, and the methodological novelty beyond this change is not sufficiently clear.

Unclear advantages of scaffold-based conditioning

  • The scaffold might be simplified, which removes fine geometric details and seems to be a drawback compared with point clouds. Furthermore, ScafPAR appears to lack the additional information required to recover these details. In terms of this part, it is unclear why scaffolds would be preferable to part-level point clouds as conditioning inputs.

Concerns about mesh visual quality and evaluation

  • Several examples contain excessive faces in flat regions and some regions also appear to contain overlapping faces. These observations of overlapping faces also raise questions about the high NME scores reported in Table 4. I assume that the large number of unnecessary edges inflates the denominator and makes the proportion of non-manifold edges appear artificially low. (e.g., the airplane with 53,820 faces, the robot with 66,339 faces in Figure 1, third and fourth examples in Figure 5).

Minor weakness

  • The Introduction does not clearly explain the methodological contributions or the proposed pipeline. These only become apparent after reading the Method section. The Introduction primarily presents module names and conceptual descriptions; it would benefit from a more concrete explanation of how the approach works.
  • The overlapping text in Figure 3 makes it difficult to follow. Improving the layout and label placement would enhance readability.
Questions for authors
  • What advantages do scaffolds offer over the part-level points as used in MeshMosaic?

  • How does the proposed method handle boundaries between parts?

  • Since the generator cannot observe the detail of original input, could it produce parts that are inconsistent with the overall geometry? Should global shape information also be provided?

  • In ScafPAR, the scaffold is embedded into a representation H. Why not directly use the latent representation before decoding instead of embedded scaffold?

  • Could the authors clarify the evaluation settings used in Table 1? MeshMosaic performs poorly, whereas, to my knowledge, it should substantially outperform TreeMeshGPT. Providing the detailed inference settings for MeshMosaic, including any deviations from its original evaluation protocol, would help explain this discrepancy and support reproducibility.

  • Could the authors discuss the apparent overlapping faces and excessive faces occur, and compare the numbers of non-manifold edges and vertices with those of other methods would be helpful to clarify my concern.

4 comments

We really appreciate the reviewer for the detailed and constructive feedback. We appreciate your recognition of independent part generation for handling longer face sequences, the use of scaffolds as conditioning inputs, and the efficiency benefits of PSAI.

We believe our response has addressed most of your concerns and we respectfully ask if you're comfortable to re-consider the rating in light of our efforts and dedication. Below, we provide detailed responses to all the comments:


Q1: FAMesh appears similar to MeshMosaic, with scaffolds replacing part-level point clouds as conditioning inputs. The methodological novelty beyond this change and the advantages of scaffold conditioning remain unclear.

Thank you for highlighting this connection. FAMesh’s key contribution is coarse–fine mesh topological traversal consistency, providing explicit connectivity and traversal guidance for fine mesh generation.

Difference from MeshMosaic. MeshMosaic only uses point clouds to constrain surface geometry and boundary triangles to connect parts. However, point clouds provide no intra-part face connectivity or traversal trajectories. Overlapping faces can therefore remain consistent with the target surface, while errors may propagate through boundary conditioning. Our experiments at high face counts reveal such overlaps, topological drift, and structural breaks (R2-Table 1 in Q4; Fig. 5).

How is the topological scaffold aligned. During per-part simplification, FAMesh tracks fine-to-coarse face assignments and orders coarse faces by the earliest traversal position of their assigned fine faces (Appendix Alg. 1; Fig. 7, right). This alignment guides topology within and across parts, mitigating error accumulation. PSAI further reduces scaffold overhead.

In Table 4, changing the traversal rule (Replace Face Traversal) has little effect on quality, whereas retaining scaffolds without alignment (w/o Traversal Consistency) substantially degrades it. Thus, coarse–fine mesh topological traversal consistency is key to effective scaffold guidance.


Q2: Can ScafPAR recover fine details and maintain global consistency from simplified scaffolds alone?

Thank you for highlighting this ambiguity. ScafPAR receives both the scaffold and global shape information Z0\mathrm{Z}_0 (Fig. 2).

The source of global shape information. As described in Section 4.1, we train a reconstruction VAE on part-level point clouds covering the entire object and use its encoded Z0\mathrm{Z}_0 to supervise DiT. At inference, DiT predicts Z0\mathrm{Z}_0 from the input image, providing global shape and geometric details for part refinement. We add the clarification for this in Appendix A.

Complementary scaffold guidance. While Z0\mathrm{Z}_0 provides geometry, the scaffold adds explicit face connectivity and traversal guidance aligned across coarse and fine meshes, which point-cloud conditioning lacks. Scaffold simplification therefore does not limit ScafPAR to coarse geometry.

Table 1 shows better geometric agreement with GT meshes than point-cloud baselines. R1-Table 1 further shows higher object-level structural success rates than competing methods and the variant without scaffolds.


Q3:How does the proposed method handle boundaries between parts?

As shown in Fig. 1 and revised Section 3.1, our mesh data represent each part as a watertight connected component with independent vertex indexing. Generated parts are directly assembled into a complete mesh. We therefore aim to maintain plausible geometric contact and reduce interpenetration, without explicit topological connections between parts. ScafPAR uses the scaffold’s part positions and shapes to coordinate contact regions, while the object-level latent Z0\mathrm{Z}_0 provides details consistent with surface geometry.

As shown in R1-Table 3, the scaffold preserves major contours and part layouts while retaining 91.15% of GT fine-mesh part adjacencies, 95.17% of scaffold adjacencies agree with the GT fine mesh, providing geometric boundary guidance for refinement.

Q4: Do excessive and overlapping faces bias NME, and how do absolute non-manifold edge/vertex counts compare across methods?

Thank you for raising this concern. We identified meshes with poor topology before receiving the reviews. While awaiting the reviews, we improved data cleaning and retrained the model, updated Figs. 1 and 5, and added examples in Appendix Fig. 9. All results are based on the latest model weights.

Fewer absolute defects at comparable face counts:

To address the NME denominator concern, we evaluate 40 test objects with comparable output face counts across methods for each object.

R2–Table 1: Mean per-mesh defect count.

MethodNon-manifold edges ↓Non-manifold vertices ↓Overlapping faces ↓Face density ↓
MeshMosaic1,769.21,307.65,619.67,619.6
MeshRipple1,337.31,684.52,472.85,755.9
FAMesh1,279.01,128.61,564.44,508.8

Each mesh’s longest bounding-box side is normalized to 1. Face density is the region’s face count divided by its normalized area. These absolute counts support mesh quality beyond the ratio-based NME metric.

Assessing redundancy in planar regions should also account for downstream accuracy:

Triangulation must respect boundaries and neighboring connectivity, and coarsening may affect numerical accuracy. We apply the Franke–Poisson test to corresponding planar regions of the same part generated by FAMesh and MeshMosaic. We progressively remove interior vertices and retriangulate, preserving each region’s geometry and inner/outer boundary vertex counts.

R2–Table 2: Progressive coarsening of planar regions with fixed boundaries.

MethodFacesFace reductionHarmonic index ↓Poisson error ↓Error change ↑
MeshMosaic5600%9.180.00241—
42025.0%8.590.00243+0.8%
35037.5%8.200.00240−0.5%
28050.0%7.570.00342+41.9%
FAMesh3200%9.210.00242—
24025.0%8.610.00309+27.6%
20037.5%8.220.00331+36.8%
16050.0%7.590.00375+54.8%

The Harmonic index measures triangle shape quality, the Poisson error measures deviation from the analytical Poisson solution. FAMesh’s original triangulation achieves comparable Poisson error with 320 faces versus MeshMosaic’s 560. Coarsening improves triangle shapes but increases FAMesh’s error by 27.6%–54.8%, whereas MeshMosaic retains comparable accuracy after a 37.5% face reduction. These results suggest less removable redundancy in FAMesh’s original triangulation and highlight the need to assess planar-region redundancy against downstream accuracy.


Q5: The Introduction lacks a concrete explanation of the pipeline and methodological contributions.

Thank you for the suggestions. We have clarified our methodological contributions and pipeline in Introduction, and adjusted Fig. 3’s layout to improve readability.


Q6: Why not use representation Hz\mathrm{H}_z?

As noted in Q2, Z0\mathrm{Z}_0 encodes point-cloud geometry, and a lightweight mapping adapts it to the autoregressive decoder input Hz\mathrm{H}_z, similar to [1]. TopoDec then decodes face connectivity and traversal trajectories part by part, forming an explicit topological scaffold to guide generation. However, using the pre-decoding latent Hz\mathrm{H}_z directly would require ScafPAR, like prior point-cloud-based methods, to infer part partitioning, layout, and contact relationships while generating fine-grained part topology from geometry alone, leading to substantial error accumulation and topological drift.

Removing the explicit scaffold, increases the part collision rate from 15.4% to 61.0% and yields a 19.68% relative decrease in human preference scores (Tables 9 and 3), demonstrating the effectiveness of explicit topological guidance.


Q7: MeshMosaic evaluation settings and performance discrepancy.

Thank you for highlighting this discrepancy. Table 1 uses same test set and native inputs. MeshMosaic uses 15 PartField hierarchical clusters, 16,384 points, temperature 0.5, a 9k-token sliding window, and 40k tokens/part. Meshes are normalized and sampled at 100k points by area for bidirectional nearest-neighbor metrics. F1/EF1 thresholds are 0.003/0.005, sharp-edge points require an absolute normal dot product <0.2 within radius 0.004.

MeshMosaic preserves parts and details and scores well in human preference (Tables 2, 10) and EF1. However, outputs show excessive face overlaps receiving repeated sampling weight and spikes that may raise HD and alter normalization, shifting and rescaling the body. These observations are consistent with higher HD/CD.

[1] EdgeRunner: Auto-regressive Auto-encoder for Efficient Mesh Generation.

Dear Reviewer,

I hope this message finds you well. As the discussion period is nearing its end, I wanted to ensure we have addressed all your concerns satisfactorily. If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and we’re eager to address any remaining issues to improve our work.

Thank you for your time and effort in reviewing our paper!

Thank you for the detailed rebuttal. The clarifications on Z₀, part boundaries, and the MeshMosaic settings are helpful, and I find the idea of aligned scaffold guidance reasonable. However, the evidence remains insufficient.

  1. The evidence that scaffold conditioning outperforms point-cloud conditioning is only indirect, as Table 1 and R2-Table 1 compare different systems and the w/o-scaffold ablation also removes explicit part partitioning. A direct comparison with part-level point-cloud conditioning under the same backbone is needed.

  2. The updated figures reflect a rendering change and do not mean the overlapping-face issue is resolved (about 1,564 per mesh in R2-Table 1). Since NME does not necessarily capture such overlaps, the high topology score in Table 4 remains questionable.

In light of the clarifications, I have raised my score to 2.

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

FAMesh generates high-face-count meshes from images in two stages. It first generates a coarse, part-aligned scaffold, then uses both the corresponding scaffold part and the full scaffold to guide detailed part generation. Fine-to-coarse face ancestry keeps the scaffold and target traversal orders aligned. The paper also introduces PSAI to compress cross-part information and reduce training cost. Experiments on an Objaverse-XL subset show improvements in geometric quality and inter-part consistency, supported by ablations and human evaluations.

Critical strengths

The main strength is the use of a part-aligned scaffold to keep detailed mesh generation structurally coherent. The ablations provide good support for this idea: without PSAI, adding TopoScaf reduces Chamfer-L1 from 0.092 to 0.025 and improves normal consistency from 0.575 to 0.924 (Table 3). Table 4 supports the importance of traversal consistency, and Table 9 shows better inter-part contact consistency and fewer collisions. Figure 8 also helps explain how future scaffold parts influence generation. The training savings from PSAI are useful, though I see them as a secondary benefit to the structural improvements.

Critical weaknesses

The effect of removing feedback from generated parts needs further evaluation. ScafPAR uses the complete scaffold to coordinate part generation, but it does not observe the actual geometry of previously generated fine parts. If a generated part deviates from its scaffold, later parts have no explicit feedback about that deviation to adjust their interfaces. Table 9 shows that scaffold guidance improves inter-part consistency, but it remains unclear how often this design still leads to substantial misalignment or other structural failures. An object-level success-rate comparison on 100 randomly sampled inputs, supported by blinded human assessment and more success and failure examples, would help clarify the practical impact of this limitation.

The results across the ablation tables need clarification. The default configuration in Table 5 reports Chamfer-L1 of 0.047, normal consistency of 0.884, and Hausdorff distance of 0.053, compared with 0.018, 0.938, and 0.027 for the full model in Tables 1 and 3. I could not find an explanation for this difference, which makes it difficult to relate the architectural ablations to the main results.

Questions for authors

Could you compare object-level success rates for FAMesh and the main baselines on the same 100 randomly sampled test inputs? The existing blinded user study could be extended to assess structural success, using clear criteria such as overall shape coherence and the absence of severe unintended fragmentation, part misplacement, or collapse. Please distinguish this from simply completing generation and producing a valid mesh file. Report successful cases out of 100 and agreement between raters, and explain how failed runs, filtering, and repeated sampling are handled. An unfiltered gallery of these outputs, with representative successes and failures shown in greater detail, would also be helpful.

Why does the default setting in Table 5 give different results from the full model in Tables 1 and 3? Were different evaluation subsets, training budgets, checkpoints, or other settings used?

Could you discuss how FAMesh relates to LATO (ICML 2026) and why it was not included in the comparison? In particular, are there differences in task settings or supported mesh scales that make a direct comparison inappropriate? This would help put FAMesh’s quality and efficiency results in context.

4 comments

We sincerely thank you for your careful review and constructive feedback. We are encouraged by your recognition of our part-aligned scaffold design and the experimental evidence supporting its benefits for geometric quality and inter-part consistency. Your encouragement is “like a lantern in the dark”, motivating us to continue refining and advancing this line of research.

Your suggestions are valuable for strengthening our evaluation of structural reliability and improving the clarity of our experimental reporting. Below, we address your questions regarding object-level structural success rates, the differences in results between Table 5 and Tables 1 and 3, and the relationship between FAMesh and LATO. We hope our responses help clarify these issues and resolve your remaining concerns.


Q1:Please assess the practical impact of ScafPAR’s potential inability to adjust subsequent part interfaces through blinded object-level comparisons across methods on 100 randomly sampled inputs, supplemented by more success and failure cases.

Thank you for the suggestion. We evaluated FAMesh's structural reliability through blind assessment on 100 randomly selected test objects.

Object-level blinded evaluation:

Each method generated one mesh per object using fixed random seeds, without reruns or cherry-picking. Six raters independently assessed anonymized results in randomized order. A mesh was considered successful if it exhibited consistent overall geometry without severe fragmentation, key-component misalignment, or structural collapse. Majority voting determined the final scores, with Fleiss’ κ measuring inter-rater agreement. Since constrained decoding guarantees parseable meshes, all failures reflect structural defects.

R1-Table 1: Object-level structural success.

MethodScaffold ↑Fine mesh ↑Fleiss’ κ ↑
MeshMosaic—350.379
Mesh-Silksong—380.493
MeshRipple—420.501
TreeMeshGPT—310.279
EdgeRunner—280.433
FAMesh (w/o TopoScaf)—260.771
FAMesh85720.802

FAMesh achieves a 72% structural success rate, outperforming the compared methods and the variant without TopoScaf.

To further examine how these sucess relate to scaffold quality, we evaluated predicted scaffolds using the same criteria and reported fine-mesh structural success rates by group.

R1-Table 2: Fine-mesh structural success rates by predicted scaffold quality.

QualitySuccesses / TotalSuccess rateFleiss’ κ
success69/8581.2%0.739
defects3/1520.0%0.696

The results show that scaffold effectively guides fine-mesh generation, while ScafPAR can partially recover from defective scaffolds.

Geometric consistency and inter-part contacts:

We compare corresponding scaffold and fine-mesh parts and assess contact relations using the adjacency metrics. Since parts are independently indexed and individually watertight, we evaluate geometric contact rather than cross-part boundary stitching.

R1-Table 3: Scaffold–fine mesh consistency and part adjacency.

SettingsIoU ↑ (%)Δc ↓ (×10⁻³)nΔc ↓ (%)nCD ↓ (×10⁻³)Adjacency precision ↑ (%)Adjacency recall ↑ (%)
GT85.781.23500.05121.667595.1791.15
Predicted79.882.46020.10141.850083.983.8

sIoU measures part surface overlap at 256³ resolution, Δc is the center offset between corresponding parts, nΔc expresses this offset as a percentage of the full scaffold’s bounding-box diagonal, nCD is the normalized surface Chamfer distance. Adjacency definitions and results follow Table 9.

These results indicate that the scaffold preserves overall shape, part layout, and contact relations, providing spatial constraints that support object-level structural consistency without explicit feedback from previously generated parts.

We add the new supplementary gallery and included representative success and failure cases in Figs. 9 and 10 of the manuscript, respectively.

Q2: Why does the default setting in Table 5 give different results from the full model in Tables 1 and 3?

We apologize for the lack of clarity. All configurations in Table 5 were trained on 50% of the training data to compare efficiency and generation quality under a consistent setting and determine the default configuration, whereas the full model in Tables 1 and 3 used the entire training set. Both use the same default architecture, number of H20 GPUs, and stopping criterion (training loss of 0.018). The evaluation set and protocol are identical, the metric differences stem from the different training data sizes. We clarify these details in the Table 5 caption and Appendix A.


Q3: Could you discuss how FAMesh relates to LATO (ICML 2026) and why it was not included in the comparison?

We thank the reviewer for pointing us to LATO. LATO employs latent flow matching to generate topological features in parallel, followed by edge prediction to recover triangles, reducing generation latency. However, its edge prediction is prone to missed connections, which can cause missing faces or disconnected components.

In contrast, FAMesh still autoregressively generates discrete sequences encoding vertex connectivity and introduces explicit topological cues through a scaffold, aiming to improve topological consistency and stability at high mesh resolutions.

Our original evaluation focused on topology drift in autoregressive generation, motivating our choice of baselines. LATO merits comparison, but its public weights cover only a mesh-to-mesh VAE whose inputs include vertex displacement fields computed from reference faces. A matched generation comparison requires the full generative weights, including VAE reconstruction in our generation quality and efficiency tables would conflate tasks.

For additional context, we evaluate the publicly released LATO VAE on the same test objects. We add Boundary Edge Ratio (BER) [1], the proportion of open boundary edges, to assess openings caused by missing faces.

R1-Table 4: Mesh quality evaluation.

MethodHD ↓CDₗ₁ ↓GT-NC ↑F1 ↑EF1 ↑BER ↓NME ↓
LATO VAE0.0640.0080.7760.3500.3780.0390.680
FAMesh0.0320.0210.8920.3310.4920.0180.019

Given complete reference meshes, LATO achieves better surface alignment (CDL1CD_{L1}, F1), but worse BER, HD, and NME indicate missing faces or discontinuities introduced during reconstruction.

[1] Mesh-RFT: Enhancing Mesh Generation via Fine-grained Reinforcement Fine-Tuning. NeurIPS 2025.

Thank you for the response. The revised paper and additional experiments ease my concerns about the generation success rate and the quantitative results. I would still appreciate a larger gallery of generated meshes in the appendix. The teaser feels repetitive, and showing more diverse objects with more complex structures would better demonstrate the method’s capabilities.

One point still needs clarification: why do the full-model results differ between Tables 1 and 3 in the revised paper? Table 1 reports Chamfer-L1 of 0.021 and normal consistency of 0.892, whereas Table 3 reports 0.018 and 0.938. Do these tables use the same model checkpoint and evaluation setup? Please clarify how they relate to the updated training data and retrained model.

For now, I maintain my rating of 3 (weak acceptance) and look forward to the other reviewers’ views.

Dear Reviewer,

We are glad to hear that the revisions and additional experiments have eased your concerns about the generation success rate and the quantitative results. Following your suggestion, we have expanded the gallery in Fig. 8 of the appendix to show a wider variety of generated objects, including examples with more complex arrangements of parts.

For Tables 1 and 3, we evaluated two different checkpoints of the full FAMesh architecture using the same test set and evaluation protocol. Table 1 reports the results after retraining on the expanded 120K-mesh dataset, with a Chamfer-L1 of 0.021 and a normal consistency of 0.892. Table 3 reports the results from the original checkpoint, with scores of 0.018 and 0.938, respectively. We have clarified which training data and checkpoint each table uses in Sections 4.1 and 4.3.

Thank you for your time and effort in reviewing our paper!

Best regards, Authors of Submission 14831

Comments on the paper

2 comments

Hi reviewers #qMDH and #6aYG, as we are approaching the end of the author rebuttal phase, could you please take a look at the authors' responses and provide your feedback asap? Thx, AC

AuthorsOpenReview ↗

General Response

We sincerely thank Reviewers rDnr (R1), qMDH (R2), and 6aYG (R3) for their thoughtful feedback. Their suggestions helped us strengthen the evaluation of structural reliability and clarify the design and experimental settings of FAMesh.

We have carefully addressed each individual comment provided by the reviewers and believe we have successfully responded to most of their concerns. In our revised manuscript, we have incorporated the suggested experiments, additional discussions, and relevant updates to further strengthen our work. Below, we summarize the core contributions of our study, the updates to our experiments, and the in-depth discussions included in our revision.


Core Contributions of Our Work

  1. Future-Informed Part Generation: Scafor predicts a scaffold that aligns with the target mesh in both part structure and face traversal order. ScafPAR then uses this scaffold to guide the generation of each part toward a topologically consistent trajectory, enabling FAMesh to produce meshes with up to 80K faces.
  2. Coarse-to-Fine Traversal Consistency: During simplification, we track which fine faces map to each coarse face. This ancestry aligns the scaffold and target traversal orders, giving ScafPAR explicit topological guidance during refinement.
  3. Efficient Cross-Part Interaction: PSAI preserves fine-grained causal attention within the current scaffold part and uses structural summaries for cross-part modeling and global scaffold guidance, yielding a 1.52× training speedup.

Updates of experimental results during Rebuttal

  • Main Text: Updated FAMesh’s results in Tables 1–2 after data cleaning and retraining (R1, R2, R3) and re-rendered Figs. 1 and 5 to make local mesh topology easier to inspect (R2, R3).
  • Appendix B: Expanded the gallery of complex meshes in Fig. 8 (R1), illustrated refinement from imperfect scaffolds in Fig. 9 (R1, R3), and compared four tokenizers in Fig. 10 (R3).
  • Individual Responses: Assessed object-level structural consistency in a blinded study of 100 objects (R1, R2) and evaluated absolute defect counts and planar-face redundancy (R2).

Updates of in-depth discussions during Rebuttal

  • Main Text: Clarified the advantages of scaffold conditioning over point clouds (R2), how the scaffold guides contact between independently generated watertight parts (R1, R2, R3), and which checkpoints were used in Tables 1 and 3 (R1).
  • Appendix A: Explained how the object-level latent preserves fine geometry (R2) and specified the 50% training subset used for Table 5 (R1).
  • Individual Responses: Clarified the scope of the LATO reconstruction comparison (R1) and documented MeshMosaic’s inference settings (R2).

We believe these additions and clarifications address the reviewers’ concerns and strengthen our manuscript. All revisions are highlighted in blue\color{blue}{\text{blue}} for ease of reference. We updated the manuscript on Nov 18, AOE time\color{red}{\text{Nov 18, AOE time}}.

We thank the reviewers again for their time and thoughtful feedback. We hope they will consider the revised manuscript favourably.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.