acceptodds
Under review as a conference paper at ICLR 2027

TrieMesh: Structured Frontier Representation for Autoregressive Artistic Mesh Generation

Abstract

Autoregressive artistic mesh generation has progressed via compact tokenization, yet each prediction step still mainly relies on sequence history to represent the current mesh state. The active frontier is an important part of this state, but is rarely represented explicitly in a structured and queryable form. We present TrieMesh, a frontier-guided autoregressive representation that maintains active frontier edges in a Trie with canonical lexicographic keys, enabling deterministic selection of the next edge for expansion. A shared autoregressive decoder switches between full-face generation when the frontier is empty and completion-vertex prediction otherwise, reusing the two known endpoints of the active edge to avoid redundant coordinate prediction. Experiments show improved geometric fidelity over recent autoregressive baselines under comparable token budgets. Beyond global generation, the same structured frontier representation naturally supports mesh hole filling, where the boundary edges of a missing region initialize the active frontier. This allows global mesh generation and local mesh editing to follow the same frontier-guided autoregressive expansion process.

Primary area:
generative models
Keywords:
Autoregressive Mesh Generation, Explicit Frontier Modeling, Trie-Based Representation, Frontier-Aware Inference

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_KTbwOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposed an autoregressive representation that maintains the evolving mesh frontier as an explicit structured state. A Trie maintains a canonical lexicographic frontier state for deterministic expansion. To avoid inconsistencies caused by re-predicting known vertices.a dual-mode inference strategy that generates a full face when the frontier is empty and otherwise reuses the two frontier vertices to predict only the new vertex.

Critical strengths

The paper addresses an interesting question about frontier representation and extends FACE with explicit frontier modeling.

Critical weaknesses
  1. The results lack a direct comparison with FACE to demonstrate the benefits of frontier modeling.
  2. The paper lacks comparisons of memory usage and generation time under the same hardware conditions.
Questions for authors
  1. What do (n) and (d) denote in Table 1?
  2. Would Figure 4 be more appropriately presented as a table?
3 comments
AuthorsOpenReview ↗

Response to Reviewer KTbw

We thank the reviewer for their constructive feedback and thoughtful questions. We address each point below.

W1: FACE Comparison and Benefits of Frontier Modeling

We would like to clarify that a direct comparison with FACE is already provided in Table 3 of our manuscript(originally labeled as “ZYX,” and now explicitly renamed to “ZYX (FACE)” in the revised manuscript). Please kindly refer to Concern 1 (“Comparison with FACE”) of our General Response for detailed explanations and the reproduced table. As shown there, under identical training and evaluation settings, TrieMesh reduces Hausdorff distance from 0.124 to 0.112 and Chamfer distance from 0.061 to 0.053 compared to ZYX(FACE), directly demonstrating the benefits of our frontier modeling and dual-mode decoding.

W2:Memory Usage and Generation Time

To further evaluate the inference efficiency of TrieMesh, we benchmarked all methods on our test set using their 7-bit versions on a single NVIDIA A100 (80 GB) GPU, ensuring a consistent quantization resolution across methods. The results are summarized below.

MethodParametersCompression Ratio ↓Avg. Time (s/mesh) ↓Avg. FacesFaces/s ↑GPU Memory (GB) ↓
MeshAnything 2350M0.46106.01,58915.33.2
BPT500M0.26102.11,14011.14.2
TreeMeshGPT420M0.2260.261,23420.44.5
MeshRipple500M1.00361.291,8775.34.5
ZYX (FACE)500M0.1131.5581,66052.74.6
Ours500M0.1234.2071,62447.24.6

As shown in the table, all methods exhibit comparable inference-time GPU memory consumption (3.2–4.6 GB), and most operate at a similar parameter scale. More importantly, we observe a general trend that methods with more compact tokenization (i.e., lower compression ratios) tend to achieve higher generation throughput. This is expected, as a lower token-to-face ratio requires fewer autoregressive decoding steps to generate the same number of faces. While actual throughput also depends on model architecture and implementation details, the compression ratio serves as an important indicator of inference efficiency in autoregressive mesh generation.

Notably, TrieMesh achieves a compression ratio of 0.12, close to the 0.11 of FACE, while maintaining a comparable generation throughput (47.2 vs. 52.7 faces/s). Moreover, both methods consume the same GPU memory (4.6 GB). These results demonstrate that our structured frontier representation and dual-mode decoding introduce little additional inference overhead while preserving the efficiency of compact autoregressive mesh generation.

Q1:What do (n) and (d) denote in Table 1

In Table 1, (O(n)) indicates that querying or deleting a specific frontier edge in a standard stack or queue requires linear time in the number of active frontier edges (n). In contrast, TrieMesh represents frontier edges using a Trie with a fixed key length of (d=6), enabling key-based lookup and deletion in (O(d)) time, which is effectively constant. We will explicitly clarify these notations in the table caption.

Q2:Would Figure 4 be more appropriately presented as a table?

Thank you for pointing out this LaTeX formatting error, which was caused by an incorrect use of the LaTeX minipage environment. We will correct this issue in the revised manuscript to ensure that the table is properly labeled and referenced as Table 2.

Thanks for the provided results, and my main concern has been addressed. I would suggest the author to add these results into the main paper. I increases my score to 3.

AuthorsOpenReview ↗

Official Comment by Authors

We sincerely thank the reviewer for the positive feedback and constructive suggestion. We are glad that the additional results have addressed your main concern. We will incorporate these results into the main paper as suggested. Thank you again for your time and thoughtful review.

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

This paper introduces an autoregressive framework that maintains frontier edges in a queryable trie. TrieMesh indexes frontier edges in a trie using the quantized coordinates of their endpoints. At each step, it selects the lexicographically smallest frontier edge and extends it with a new face. The selected edge is removed, while each of the other two edges is deleted if it already exists in the trie or inserted otherwise. Through these insertion, lookup, and deletion operations, TrieMesh progressively expands the mesh while preserving manifoldness. This representation enables mesh generation with fewer tokens than existing methods while producing stable results.

Critical strengths
  1. Propose an efficient frontier representation that orders frontier edges lexicographically, reducing the token count—a challenge in existing mesh generation methods.

  2. Design a dual-mode inference scheme tailored to this representation, enabling efficient and stable inference compared with methods such as FACE that generate all vertices explicitly.

  3. Demonstrate the applicability of the method to mesh hole filling, highlighting its potential for further applications.

Critical weaknesses

Ineffienct of reasoning for next face

The autoregressive face decoder generates faces without consulting the trie frontier, requiring it to implicitly infer which frontier edges remain available. Explicitly conditioning the selection of the next face on the remaining frontier edges could substantially reduce the set of possible choices and potentially improve generation stability. In the current architecture, however, the decoder must learn to track and reason about the frontier implicitly.

7-bit discretized space

The paper uses a 7-bit discretized space as its target representation, which limits its ability to reliably capture fine geometric detail. Simply increasing the resolution to 9-bits could make generation less stable, yet the paper does not propose a method to address this challenge. This raises a significant concern about whether the proposed architecture, and thus the paper’s main contribution, can generalize beyond the 7-bit setting.

Minor weakness

The introduction would benefit from a brief explanation of what “lexicographic” means in this context. Even a short, intuitive description would help readers follow the introduction as a whole.

Questions for authors
  • Without explicit access to the frontier, the Transformer decoder must select the next face from a very large space of possibilities, which could lead to unstable generation. Have the authors considered conditioning the decoder on the remaining frontier edges to constrain the candidate faces and improve stability?

  • To my knowledge, causal MLP–based approaches become highly unstable and produce incorrect predictions when using 9-bit discretization. Does the proposed method face similar issues, and was the choice of 7-bit discretization motivated by these concerns? If so, how do the authors address the loss of geometric detail at this resolution? Is this a fundamental limitation of the current architecture?

5 comments
AuthorsOpenReview ↗

Response to Reviewer fNGS

We thank the reviewer for their constructive feedback and thoughtful questions. We address each point below.

W1:Inefficient Reasoning for the Next Face

We would like to clarify that TrieMesh does not require the Transformer decoder to implicitly track the active frontier or select the next frontier edge, as these operations are explicitly handled by the Trie during inference.

Specifically, during training, face sequences are pre-serialized according to the Trie-based traversal order. During inference, the Trie explicitly maintains the active frontier and deterministically selects the next edge for expansion. Accordingly, the decoder operates in two modes: full-face generation when the Trie is empty, and completion-vertex prediction when the Trie is non-empty, where the two endpoints of the selected frontier edge are reused and only the third vertex is predicted.

As the reviewer noted, conditioning face generation on the frontier reduces the prediction space and can improve generation stability, which is a key motivation behind TrieMesh. Our design achieves this through explicit frontier management and dual-mode decoding, without introducing additional learnable parameters for frontier tracking or selection. The effectiveness of this design is further supported by the ablation results in Table 3, where our dual-mode approach achieves lower Hausdorff and Chamfer distances than both the ZYX(FACE) baseline and the full 9-slot variant.

W2 and Q2: 7-bit Discretization and Resolution Scaling

We thank the reviewer for raising this concern.

Why we report results at 7-bit. Our choice of 7-bit quantization is motivated by experimental comparability rather than an architectural limitation. TrieMesh is trained on approximately 100K meshes, following a setting comparable to existing autoregressive mesh generation methods, in particular the 7-bit base model of FACE. As our primary contribution lies in frontier-guided generation rather than resolution scaling, this setting allows us to isolate the effect of the proposed frontier representation and dual-mode decoding.

On higher-resolution generation. Scaling to 9-bit is governed by several factors, including training data scale, model capacity, and input point cloud density, rather than by quantization alone. FACE reports 10-bit quantization in Section 4.6 of their paper, using a 1.2B-parameter model trained on 380K high-quality meshes with 65,536 sampled input points. This indicates that higher-resolution generation is reachable when the model and data budget are scaled accordingly.

Is this a fundamental limitation? We do not think so. TrieMesh introduces no quantization-specific restrictions: its Trie-based frontier management, deterministic edge selection, and dual-mode decoding operate on mesh topology and are not tied to 7-bit coordinates. Moving to a higher resolution only enlarges the per-axis vocabulary and the output head, while the mechanism itself stays the same, so no part of our method needs to change. Together with the scaling evidence above, we expect TrieMesh to extend to higher resolutions under comparable training conditions, and we will add this discussion to the revised manuscript.

W3: Explanation of“lexicographic”

Thank you for this helpful suggestion. In the context of autoregressive mesh generation, lexicographic ordering (specifically, ZYX order) refers to a dictionary-style sorting of 3D coordinates: vertices are ordered first by their zz-coordinate, breaking ties by yy, and then by xx. Each triangle face is represented by its three vertices sorted in this ZYX order, and all faces in a mesh are then serialized by comparing their sorted vertex tuples in the same dictionary manner (from the lowest to the highest coordinates). We will add added this explanation to the Introduction of the revised manuscript to improve clarity.

Q1: Frontier Conditioning for Next-Face Prediction

We agree that explicitly conditioning face generation on the frontier can reduce the prediction space and potentially improve generation stability. However, encoding all active frontier edges as additional Transformer tokens would increase the context length, attention computation, and KV cache memory, particularly as the frontier size varies dynamically (see Figure 11).

Instead, TrieMesh maintains the complete frontier explicitly using a Trie, while conditioning the CausalMLP only on the selected frontier edge (see Lines 231–251). This design avoids the overhead of encoding the entire frontier into the Transformer, while retaining the benefits of frontier-guided generation without introducing additional learnable parameters for frontier tracking or selection.

The authors' justification for using 7-bit quantization is insufficient. The same ground-truth mesh dataset could, in principle, be quantized at 9-bit or 10-bit resolution without requiring additional training data. Furthermore, FACE's ability to support 10-bit quantization does not establish that TrieMesh can achieve similarly stable generation, as the two methods employ different prediction architectures. In particular, the behavior of TrieMesh's causal MLP-based prediction mechanism under finer discretization remains unverified. Therefore, I strongly encourage the authors to conduct at least a 9-bit experiment using their existing training dataset to demonstrate the stability and feasibility of higher-resolution generation. If this is not feasible, the authors should explicitly acknowledge the lack of high-resolution evaluation as a limitation rather than claiming that their architecture is not fundamentally constrained by quantization resolution.

AuthorsOpenReview ↗

Further Response to Reviewer fNGS: 7-bit vs. 9-bit Quantization

Thank you for your reply and suggestion. We compared 7-bit and 9-bit quantization while keeping the model (a 24-layer decoder-only Transformer with a hidden dimension of 1024), the training data (our existing 100K meshes with fewer than 5,000 faces), and the training schedule (100K steps) unchanged. We evaluate two decoding strategies: Full 9-slot decoding generates all 9 tokens of each face from scratch, while our dual-mode decoding, enabled by the Trie-guided face ordering, reuses the 6 tokens of the selected frontier edge when the Trie is non-empty and generates only the 3 tokens of the remaining vertex. This is the core innovation that distinguishes TrieMesh from FACE. The results on our test sets (see lines 373–374) are shown below:

Objaverse║Toys4K
Bits \ HD/CD ↓Dual-mode (Ours)Full 9-slot║Dual-mode (Ours)Full 9-slot
7-bit0.094 / 0.0360.134 / 0.041║0.064 / 0.0270.087 / 0.038
9-bit0.153 / 0.0540.195 / 0.077║0.132 / 0.0440.184 / 0.071
Performance drop (7-bit → 9-bit) ↓0.059 / 0.0180.061 / 0.036║0.068 / 0.0170.097 / 0.033

We do observe a performance drop from 7-bit to 9-bit. However, dual-mode decoding consistently outperforms Full 9-slot decoding at both resolutions, and suffers a smaller drop on all metrics and both datasets (e.g., CD increases by only 0.018 / 0.017 vs. 0.036 / 0.033 on Objaverse / Toys4K). We attribute this robustness to the reduced number of tokens predicted per face, which mitigates error accumulation at higher resolutions.

Moreover, we do not think this drop should be attributed to a limitation of the CausalMLP architecture. Evidence from both mesh generation and LLM literature further supports this view. In the mesh generation domain, BPT [1] shows that finer discretization is achieved by enlarging the vocabulary (Supp. "Finer Discretization of BPT"), and that an overly large vocabulary makes prediction harder and degrades generation quality under a fixed quantization resolution (Sec. 5.4, Table 3). BPT further shows that training on meshes with more faces significantly improves generation performance (Sec. 5.4, Fig. 8), whereas our training meshes contain fewer than 5K faces. In the LLM domain, Tao et al. [2] show that the optimal vocabulary size grows with the compute budget, i.e., larger vocabularies require larger models and more data (Abstract and Fig. 1), and that larger vocabularies are prone to under-fitting rare tokens in the data-constrained regime (Sec. 1). Moving from 7-bit to 9-bit enlarges the coordinate vocabulary from 128 to 512, while our model capacity, training data (meshes with fewer than 5K faces), and training schedule (100K steps) remain unchanged. Therefore, the observed drop is more plausibly explained by insufficient model and data scale for the larger vocabulary, rather than an inherent limitation of the architecture.

[1] H. Weng, Z. Zhao, B. Lei, et al. Scaling Mesh Generation via Compressive Tokenization. CVPR, 2025.

[2] C. Tao, Q. Liu, L. Dou, et al. Scaling Laws with Vocabulary: Larger Models Deserve Larger Vocabularies. NeurIPS, 2024.

Reviewer_Ao7kOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TrieMesh, a frontier-guided autoregressive mesh generation framework that uses a Trie-based representation to explicitly manage active frontier edges and efficiently generate new faces by reusing known vertices. Experiments demonstrate improved geometric fidelity over existing autoregressive methods, with further applications to mesh hole filling. Overall, the motivation behind the proposed design and its key distinctions from existing methods are not sufficiently clear.

Critical strengths

To the best of my knowledge, a frontier-guided autoregressive mesh generation framework employing a Trie-based representation has not been explored in prior work.

Critical weaknesses
  • Limited novelty and missing comparison with FACE. This paper builds heavily on FACE, and the proposed method appears to offer only incremental improvements. More importantly, the experimental evaluation lacks a direct comparison with FACE, making it difficult to assess the actual benefits of the proposed method over its primary baseline.

  • Unclear motivation for artistic mesh generation. Although the paper focuses on artistic mesh generation, it is unclear how the proposed design specifically addresses the unique characteristics or requirements of artist-created meshes. The overall framework appears to be applicable to general mesh generation, and the authors should clarify what makes the proposed method particularly suitable for artistic meshes.

Questions for authors

Please explain the concerns in Critical Weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer Ao7k

We thank the reviewer for their constructive feedback and thoughtful questions. We address each point below.

W1:Limited novelty and missing comparison with FACE

Thank you for raising these concerns. Please see Concerns 1 and 2 in our General Response. In short: Table 3 already contains a matched FACE-style baseline (to be labeled “ZYX (FACE)”); we will also include this comparison in the main results table; and the ablation shows that the quality gain over FACE-style decoding is driven by frontier-conditioned dual-mode prediction, together with explicit frontier management and a natural extension to mesh hole filling.

W2: Motivation for Artistic Mesh Generation

We thank the reviewer for raising this question. We agree that our proposed autoregressive framework is not limited to artistic mesh generation and is also applicable to general mesh generation.

The term “artistic mesh generation” was not introduced by our work, but follows the terminology established in prior mesh generation studies, such as MeshGPT, MeshAnything, TreeMeshGPT, and their successors. In this line of research, artist-created meshes are typically distinguished from dense meshes produced by conventional isosurface extraction (e.g. marching cubes), with an emphasis on compact and explicitly structured triangulations. We therefore adopt this terminology to maintain consistency with prior work.

While acknowledging the broader applicability of our method, we retain the term “artistic mesh generation” to reflect this established task setting and distinguish it from conventional surface extraction approaches, rather than to imply that TrieMesh is exclusively applicable to artist-created meshes.

My main concern has been addressed. I changed my score to 3.

AuthorsOpenReview ↗

Official Comment by Authors

Thank you for your positive feedback. We are glad that your main concern has been addressed. We also appreciate your comments on the broader use of our representation in general mesh generation. We found this perspective very helpful, and it gives us a useful direction for future work.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all reviewers for their thoughtful and constructive feedback. In this general response, we address two main concerns raised by the reviewers: (1) the comparison with FACE, and (2) the motivation and key contributions of TrieMesh.

(Concern 1) Comparison with FACE

We would like to clarify that a comparison with FACE is already included in Table 3 as part of our ablation study, where the “ZYX” setting corresponds to our reimplementation of FACE.

Specifically, FACE adopts a face-level autoregressive architecture with a naive ZYX lexicographic ordering of triangle faces. The “ZYX” setting in our ablation study follows this baseline configuration. Since FACE's official code and pretrained weights were not publicly available, we reimplemented the method following the FACE paper, and trained and evaluated it under the same experimental settings (see Lines 406–414). This comparison was originally included in the ablation study (Table 3), as it was mainly intended to evaluate the improvements of TrieMesh over the naive ZYX ordering. To avoid ambiguity, we will rename “ZYX” to “ZYX (FACE)” and move this comparison to the main comparison in the revised manuscript, together with a more detailed description of how FACE is reimplemented. The updated Table 3 is shown below.

SettingHausdorff ↓Chamfer ↓
ZYX(FACE)0.1240.061
Ours (Full 9-slot)0.1290.066
Ours (Dual-mode)0.1120.053

The results show that TrieMesh, which combines a structured frontier representation with dual-mode decoding, outperforms both the ZYX (FACE) baseline and the full 9-slot variant. These results confirm that Trie-guided inference indeed leads to improved generation quality.

(Concern 2) Motivation and Key Contributions

As Reviewer fNGS noted, conditioning face generation on the frontier can narrow the prediction space and improve stability, which is a core motivation behind TrieMesh. To this end, we introduce a Trie-guided face order coupled with a dual-mode decoding strategy, enabling frontier-guided autoregressive mesh generation while maintaining a low compression ratio of 0.12.

Furthermore, beyond FACE, TrieMesh naturally extends to mesh hole filling (see Section 5 and Appendix F), where the boundary edges of a missing region initialize the Trie, enabling local mesh completion through the same autoregressive frontier expansion mechanism.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.