acceptodds
Under review as a conference paper at ICLR 2027

MeshCarve: Artisan Mesh Generation with Flow Matching in Compact Latent Spaces

Abstract

Prior artisan mesh generation works largely predict face tokens autoregressively, which makes inference slow. Recent methods instead flow match continuous latents built by Variational AutoEncoders (VAEs), but reconstruction quality drops significantly when geometry and topology are jointly encoded, and further when the latent space is compressed. We present MeshCarve, a flow matching method that generates entirely in compact latent spaces, generating vertex positions and edge connections separately and sidestepping the difficulty of a joint compact latent. To shorten the token sequence, we propose a hierarchical sparse transformer backbone, instantiated as VertexVAE and EdgeVAE. Instead of encoding fields over the surface voxels, both VAEs anchor on discrete vertices in their latent spaces, which drastically reduces the token sequence length, and our spatial-aware compression shortens it further without costing reconstruction. VertexVAE directly encodes vertex occupancy. For connectivity, we propose vertex-link encoding, which turns arbitrary connectivity between vertices into fixed-length continuous per-vertex embeddings and recovers complex artistic topology faithfully. MeshCarve combines these VAEs with an anchor generator and flow matches on the shortened token sequences. It shows advantages over state-of-the-art autoregressive and flow matching methods on Objaverse and generalizes to Toys4K. To the best of our knowledge, it is among the first artisan mesh generation methods whose every generative stage runs in a spatially compressed latent, with a token sequence only a fraction of the most compressed previous autoregressive and flow matching works.

Primary area:
generative models
Keywords:
Artisan Mesh Generation, Flow Matching, Variational Autoencoder, Latent Representation

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

This paper introduces MeshCarve, an artistic mesh generation framework from input point cloud. MeshCarve separately generates vertex and edge both in compact latent spaces, supported by its proposed VertexVAE and EdgeVAE. Spatial-aware compression and vertex-link encoding improve the latent compactness and thus diffusability / scalability. Experiments demonstrate competitive surface fidelity and strong connectivity generation when reference vertices are provided, compared with recent SOTA methods including meshflow, LATO.2.

Critical strengths
  1. A compact representation design for diffusion-based mesh generation. Although previous methods have already explored separate vertex-edge latent diffusion system or edge connectivity encoding & decoding algorithm, MeshCarve pushes it into a more compact latent space for both vertex and edge, especially for edge latents. Also, vertex anchoring, permutation-invariant neighborhood encoding, and spatially aware compression form a well-motivated approach to reducing generative sequence length.
  2. Both qualitative and quantitative results demonstrate clear improvements over strong baselines such as LATO.2. Detailed ablations further show the effects of different design choices.
  3. Improved adherence to the requested vertex budget, as shown in supplementary Table 5. This also reflects the flexibility of compact latent generation to some extent. This is a useful practical capability for applications where mesh complexity must be controlled.
Critical weaknesses
  1. MeshCarve requires point cloud as input, and does not demonstrate the image-to-mesh generation ability. Although point-to-mesh is already useful, I am curious whether the whole pipeline could effectively handle image input, especially the single-forward coarse anchor generation model. This coarse occupancy accuracy is generally critical to the downstream models.
  2. It would be better to show different generated meshes under various vertex budget for the same input, to demonstrate the effectiveness of budget control.
  3. The contribution of compression to generation quality is not sufficiently isolated. Appendix I reports that the released LATO.2 vertex and topology flows contain approximately 162M and 221M parameters, whereas MeshCarve uses 1.27B and 2.37B parameters. Consequently, its generation gains may reflect both the proposed representation and substantially greater model capacity. The experiments do not establish how much improvement compression provides at a fixed model size or compute budget.
Questions for authors
  1. What fraction of the topological defects is attributable to face recovery? Could the authors reconstruct faces from ground-truth vertices and ground-truth adjacency using the current three-cycle enumeration procedure, and report face precision/recall and non-manifoldness? How do the authors explain the substantial non-manifold edge percentages that remain with reference vertices?
  2. In Table 3a, the reconstruction performance of LATO.2 V-VAE is very poor. Yet the performance of end-to-end generation of LATO.2 in Table 2 is very close to MeshCarve (even better in HD). How should this be interpreted?
2 comments

Thank you for the positive assessments and the constructive questions. With all weights unchanged, we now sample the vertex flow from a start-noise temperature of 0.7 instead of 1.0 with 25 steps for faster and better generation results. The limitations that were pointed out are discussed in the revised abstract and main text, and we also answer each point below for your convenience.

1. Image input (W1)

Thank you for raising this. We agree that image to 3D generation is a valuable direction. At this moment MeshCarve cannot take an image as input, and we state this limitation in the revised main text. MeshCarve is conditioned on a point cloud by design and focuses more on designing the VAE that achieves better compression to save training/inference cost. Though interesting, image-to-mesh generation is beyond the scope of this submission, so we leave it to future work.

2. Several vertex budgets for one input (W2)

Thank you for the suggestion. A new figure in the revised PDF shows four Objaverse inputs, each generated at vertex budget scales of 0.5, 0.75, 1.0, 1.25 and 1.5. Over the 185 Objaverse meshes, the number of generated vertices rises monotonically with the scale, from 0.61 to 1.34 times the count at scale 1.0.

3. Compression and model capacity (W3)

Thank you for raising this point, and we ran a new experiment for it. On speed, our inference latency is directly comparable to LATO.2, 5.76 s against 4.99 s per mesh on one H200, despite the larger flow models.

We designed the VAE to have strong compression ability so that the cost of training larger flow matching models can be significantly reduced. We have studied the scaling of the edge flow model during this rebuttal period and the results are shown below. At 226M size, directly comparable to LATO.2's, we train our model with the same recipe as our reported 2B model, it mostly surpasses LATO.2's T-FLOW model on both evaluation sets and lags behind the 2B model. This suggests that 1) our VAE design does contribute to the performance uplift, and 2) scaling the flow matching model is useful and also improves the generation quality.

Toys4K / Obj.Edge flow F1 ↑NM ↓ (%)
LATO.2 (221M T-FLOW).729 / .67321.1 / 32.4
MeshCarve (226M edge flow).772 / .76919.0 / 23.9
MeshCarve (2B edge flow).795 / .82011.7 / 14.9

Both columns are measured with the edge flow given the reference vertices. The full metrics of the 226M model are in Appendix L of the revised PDF.

4. Topology error from face recovery (Q1)

We ran the 3-cycle enumeration on the reference vertex voxels at 512³ with their reference edges:

FacesFace PFace RNM
Objaverse: reference faces at 512³––2.1
Objaverse: 3-cycles of reference edges.9831.0004.7
Toys4K: reference faces at 512³––0.6
Toys4K: 3-cycles of reference edges.9891.0002.5

The enumeration recovers every reference face and adds 2.6 and 1.9 points of non-manifold edges. It remains an interesting direction to avoid these losses introduced by 3-cycle and we leave it for future work. With reference vertices, our non-manifold edge ratio is already lower than that of concurrent work. Flow matching models lead in surface fidelity and in handling complex shapes, but their topology, including that of current commercial models, is still not perfect. Our future work will focus on improving the topology quality.

5. LATO.2 end to end despite its V-VAE (Q2)

The vertex F1 of Table 3a requires an exact match of each vertex voxel, which LATO.2 also uses in their work. We found that their released V-VAE reconstruction performance decreases beyond the resolution of 256. Regarding their end-to-end performance, metrics such as CD and HD are more sensitive to surface fidelity and less sensitive to the exact location of individual vertices. In general, for VAEs, higher reconstruction performance is preferred.

Thank you for the detailed rebuttal. My major concerns are largely resolved. I will keep my positive rating.

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

The paper follows a three-stage pipeline: anchor generation, vertex generation, and edge prediction. Its main selling point is that it uses only 0.13 latent tokens per triangle on average, compared with 0.52 for MeshFlow. The advantage is a higher compression ratio, but the downside is underwhelming generation quality. For mesh generation, increasing compression at the expense of quality offers limited value. Nevertheless, the overall pipeline makes sense.

Critical strengths

higher compression ratio

Critical weaknesses

For end-to-end generation, non-manifold edges account for 27.7% of edges on Objaverse and 17.5% on Toys4K.

Questions for authors

The paper is missing a citation to iFlame.

1 comment

Thank you for the review and for recognizing the compression ratio of our representation. We appreciate your comment on generation quality, and we have improved it since the submission. With all weights unchanged, we now sample the vertex flow from a start-noise temperature of 0.7 instead of 1.0 with 25 steps. Non-manifold edges drop from 27.7% to 20.8% on Objaverse and from 17.5% to 15.7% on Toys4K, and CD-L2 on Objaverse improves from 64.9 to 59.7 (×10⁻⁴).

The remaining limitations are discussed in the revised abstract and main text. Thank you also for pointing us to iFlame (Wang et al., 2025), which the revised PDF now cites in the related work on autoregressive methods.

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

MeshCarve generates an artist-style mesh from a dense input point cloud and a requested vertex budget. It predicts a coarse vertex anchor, then separately flow-matches vertex positions and edge connectivity in compressed sparse latents. VertexVAE represents vertex occupancy; EdgeVAE uses an order-invariant, fixed-width sum of random Fourier features over each vertex’s neighbor offsets. Triangular faces are recovered as 3-cycles of the predicted adjacency graph. The authors report 0.13 tokens per face, training on 558K Objaverse meshes, and evaluation on 185 held-out Objaverse meshes and 500 Toys4K meshes.

Critical strengths
  • The vertex-anchored sparse representation is a concrete response to long sequences. The order-free vertex-link encoding avoids a fixed maximum degree, and its ablation is consequential: EdgeVAE F1 falls from 0.96965 to 0.70793 without the encoding and to 0.88111 with mean pooling and broadcast (Table 4, p. 9).
  • The paper includes controlled component studies and separates a conditional edge-flow evaluation from end-to-end generation. Given reference vertices, the edge flow beats LATO.2 T-Flow on all four reported surface metrics on both datasets (Table 2, p. 7).
Critical weaknesses

End-to-end topology is far from that of artist meshes, and the main comparison does not measure it.

  • The paper's target is artisan topology, but its main-text evidence for state-of-the-art performance (Table 2) consists only of surface metrics, which cannot see connectivity. Appendix D (Table 6) reports statistics of the generated graphs: (i) End to end, 27.7% of MeshCarve's edges are non-manifold on Objaverse and 17.5% on Toys4K, against 0.0% and 0.2% for the artist references and under 7% for every autoregressive baseline except FastMesh. (ii) On Objaverse, 16.8% of its edges are boundary edges, against 5.6% for the references. (iii) Even given the reference vertices, the edge flow produces 14.9% (Objaverse) and 11.7% (Toys4K) non-manifold edges.
  • Appendix B explains part of the cause: every 3-cycle of the predicted edge graph is emitted as a face, with no face classifier and no manifoldness constraint. Still, a mesh in which over a quarter of the edges are non-manifold is not directly usable for the editing, deformation and texture mapping that the introduction gives as motivation. The claimed advantage over the state of the art therefore holds for surface fidelity, not for topology, where the autoregressive methods remain closer to artist meshes. It would be useful to acknowledge this limitation in the main text and abstract.
  • The end-to-end results support a narrower conclusion than the broad advantage claim. On Objaverse MeshCarve has the lowest reported Chamfer distances among the listed methods, but a higher Hausdorff distance than MeshFlow and LATO.2. On Toys4K, its Chamfer results are effectively tied with LATO.2 (CD-L2 0.00560 vs. 0.00562; CD-L1 0.00814 vs. 0.00820), while its Hausdorff distance is higher (0.0666 vs. 0.0476; MeshFlow reports 0.0355; Table 2, p. 7). The paper gives no uncertainty estimates for these 185/500-shape evaluations. Also, 0.13 tokens per face measures sequence length, not end-to-end latency, throughput, FLOPs, or memory; the pipeline has an anchor model and two large flows, with the flow models reported at 1.27B and 2.37B parameters (§4, p. 7).

Several comparisons and ablations use protocols that favor the proposed method or cannot isolate the source of its gains.

  • Vertex F1 is biased toward MeshCarve's representation. Table 3a requires an exact voxel match at 512³. MeshFlow's regressed positions are quantized to 512³ and LATO.2's are scored before its sub-voxel offset (Appendix C), which explains MeshFlow's F1 of .017. The edge-flow F1, by contrast, snaps LATO.2's vertices to the nearest reference voxel within one cell. With inconsistent tolerances, the vertex comparison mostly tests whether a method shares MeshCarve's voxel representation.
  • Most ablations were fine-tuned rather than trained from scratch. The mean-pooling, w/o vertex-link encoding and w/o RoPE variants were fine-tuned from the deployed VAE for 25K steps (Appendix C). Removing a component that a trained model depends on and then fine-tuning briefly biases the result against the variant, so the .409 F1 for w/o RoPE likely overstates the effect. The one from-scratch comparison against vertex-link encoding, the slotted encoding, is within 0.8 F1 points (.962 vs. .970), so the measured benefit over the prior design is small.
  • Compression and model scale are confounded. The authors attribute the edge-flow lead mostly to compression, but MeshCarve's flows have 1.27B and 2.37B parameters, against 162M and 221M for LATO.2 (Appendix I). Appendix E isolates sequence length, not generation quality. The EdgeVAE also reconstructs edges worse than LATO.2's (F1 .973/.962 vs. .999/.997, Table 3a), which contradicts the abstract's claim that compression comes without a reconstruction cost. A controlled experiment, such as using a less-compressed edge latent with a similarly sized flow model, would help clarify the contribution of compression.
  • Small test set without variance. The main Objaverse comparison uses 185 meshes, with no confidence intervals or seeds, and no description of how the set was selected or deduplicated against the 558K training meshes and the baselines' training data. Some matched subsets are very small (15 Objaverse meshes for MeshAnything, Table 7).
Questions for authors
  • Topology quality: The generated meshes still have a relatively high percentage of non-manifold edges. Could you include these structural statistics in the main paper and report how much they improve after applying a simple face filter or using the edge flow with generated vertices? Does this also affect the surface quality metrics?
  • Vertex F1 evaluation: Could you evaluate vertex F1 using the same matching tolerance for all methods, rather than requiring an exact voxel match? Also, could you clarify whether the LATO.2 V-VAE was reproduced or evaluated using the released weights? The main text and Appendix C seem to describe different settings.
  • Ablation study: Could you train the w/o RoPE and w/o vertex-link encoding variants from scratch, as done for the slotted encoding variant? Also, what representation is used in place of vertex-link encoding in this ablation?
  • Effect of compression: How much of the improvement in edge generation comes from the proposed representation, and how much could be due to the larger flow model? Could you provide a comparison with a less-compressed latent or a smaller model, while keeping the model size or compute budget comparable?
  • Inference time: Could you report inference time for LATO.2, MeshFlow and at least one autoregressive baseline on the same GPU? It would also be useful to see how the inference time of MeshCarve changes for meshes larger than those included in the current timing results.
  • Vertex budget: How sensitive is the method to the choice of vertex budget? In particular, what happens when the budget differs from the reference vertex count by around 25%, or when the budget is predicted rather than provided from the reference mesh?
  • Test-set selection: Could you clarify how the 185 Objaverse test meshes were selected and whether they were checked for duplicates or overlap with the training data of MeshCarve and the baseline methods?
  • Anchor-generator ablation: The flow-based anchor generator appears to have been trained for the same number of steps as the proposed anchor generator. Since flow-based models may require a different training schedule, does the performance gap remain when the baseline is trained for longer?
  • Input point cloud: Appendix B mentions that the anchor generator uses point samples from a dual-grid remesh. Could you clarify whether the frozen shape encoder also receives points sampled from this remesh or from the original artist mesh? This distinction matters because the latter could retain information about the original faces and vertex locations.
2 comments

We appreciate the detailed review. With weights unchanged, we now sample the vertex flow from a start-noise temperature of 0.7 instead of 1.0 with 25 steps for faster and better generation results. We thank you for pointing out the limitations and these are discussed in the revised abstract and main text.

1. Topology and surface fidelity (W1, Q1)

Flow matching methods achieve superior surface fidelity, speed, and ability to handle complex shapes compared to prior AR methods, though we agree that they, including commercial options, still have limitations in generating perfect topology. Our future work will focus on improving the topology quality. Please see the updated comparison against other flow matching baselines below. Regarding HD, qualitatively MeshFlow reaches a better HD with low-poly meshes but significantly underperforms on CD. LATO.2’s surplus faces likely helped cover holes but led to more non-manifold faces. For brevity, the full CD-L1 results and the structural statistics (B, M and NM) are in the main comparison table of the revised PDF.

ObjaverseCD-L2 (×10⁻⁴)HD (×10⁻²)NC (×10⁻²)BMNM
MeshFlow82.25.2581.406.466.427.3
LATO.269.27.5782.476.548.445.1
Ours (paper)64.97.8985.2916.855.427.7
Ours (new)59.76.8386.2712.067.220.8
Toys4K
MeshFlow68.43.6388.112.782.414.9
LATO.257.54.7188.433.461.934.6
Ours (paper)56.06.6690.6917.265.317.5
Ours (new)54.65.9891.0513.171.215.7
max SD (4 seeds)1.60.330.231.23.02.7

A simple face filter is an interesting idea. We implement it by keeping at most two faces per edge (largest minimum angle first). More complex post-processing is left to future work.

New infer. configNMCD-L2 (×10⁻⁴)HD (×10⁻²)NC (×10⁻²)
Objaverse, before20.859.76.8386.27
Objaverse, after0.060.36.9186.68
Toys4K, before15.754.65.9891.05
Toys4K, after0.055.15.7691.27

Regarding comparison to LATO.2. Per reviewer guideline, LATO.2 is a concurrent ICLR 2027 submission, which authors are not required to compare against. It was nevertheless included in our work. The uncertainty is given in the revised table above, for brevity, only the max standard deviation is given. The revised PDF includes the mean of multiple draws on the evaluation datasets for the flow matching based methods. Appendix K of the revised PDF reports the standard deviation of MeshFlow, LATO.2 and MeshCarve. The camera-ready version will include it for the remaining AR baselines.

2. Efficiency and compression (W1, W2, Q4, Q5)

Speed with the new inference config is below. Our inference latency is directly comparable to LATO.2.

One H200, bs 1, 24 meshesMeshFlowLATO.2OursFastMeshMeshAnythingMeshAnything V2TreeMeshGPTBPTDeepMesh
s / mesh2.434.995.7622.243.050.370.5147.8390.3

From 500 to 20K reference vertices, the time of MeshCarve grows from 5.6 s to 7.5 s; that of LATO.2 grows from 2.8 s to 7.3 s by 4.6K vertices, and its T-FLOW is limited to 10K vertices.

MeshCarve takes about 420 TFLOPs per mesh on the same 24 meshes (anchor generator 14.5, vertex flow 255, edge flow 150). The peak VRAM usage of MeshCarve at inference is about 8 GB.

For the relationship between the VAE and flow matching backbone. We designed the VAE to have strong compression ability so that the cost of training larger flow matching models can be significantly reduced. We have studied the scaling of the edge flow model and the results are shown below. At 226M size, directly comparable to LATO.2's, we train our model with the same recipe as our reported 2B model, it mostly surpasses LATO.2's T-FLOW model on both evaluation sets and lags behind the 2B model. This suggests that 1) the edge VAE contributes to the performance uplift, and 2) scaling the flow matching model is useful and also improves the generation quality.

Toys4K / Obj.Edge flow F1 ↑NM ↓ (%)
LATO.2 (221M T-FLOW).729 / .67321.1 / 32.4
MeshCarve (226M edge flow).772 / .76919.0 / 23.9
MeshCarve (2B edge flow).795 / .82011.7 / 14.9

Both columns are measured with the edge flow given the reference vertices. The full metrics of the 226M model are in Appendix L of the revised PDF.

Regarding the reconstruction claim in the abstract. It is made against the prior peer-reviewed work MeshFlow. LATO.2 is a concurrent ICLR 2027 submission, so the claim is not targeted at LATO.2.

3. Evaluation protocol and ablations (W2, Q2, Q3)

The reported vertex F1 is not biased towards any methods. It is the same metrics that LATO.2 used, i.e., the F1 metrics of vertex octree subdivision process. The sub-voxel offset process as LATO.2 suggested, only operates within a 1024 resolution voxel and does not decrease nor improve the vertex F1. The one-cell snap in the edge-flow F1 favors LATO.2, not our method, as stated in Appendix C of our submission. In LATO.2’s arXiv submission, page 12, Table 4, we quote “Removing the offset head only affects continuous sub-voxel localization, degrading CD and HD but leaving voxel-level occupancy metrics unchanged.”, which directly clarifies that the sub-voxel offset should not affect the vertex F1.

The word “reproduce” in our paper was inaccurate. We only ran the LATO.2 V-VAE from its released code and weights. We did not retrain nor modify their model. The revised PDF corrects this wording. Regarding MeshFlow, regression does not excuse a low vertex F1. Its VAE fails to reconstruct the vertex positions accurately, even under the more tolerant metrics below.

Vertex reconstruction (Toys4K / Objaverse)Ours (generates at 512³)LATO.2 V-VAE (released weights)MeshFlow VAE
F1 ↑, exact cell at 256³.9995 / .9989.960 / .912.085 / .091
F1 ↑, exact cell at 512³.9984 / .9959.375 / .343.017 / .020
F1 ↑, exact cell at 1024³–.145 / .117.003 / .003
F1 ↑, 512³ with 1-cell tolerance.9990 / .9978.557 / .525.059 / .062
F1 ↑, 512³ with 2-cell tolerance (= 256³ with 1-cell).9997 / .9994.965 / .948.271 / .281
F1 ↑, 256³ with 2-cell tolerance.9999 / .9997.999 / .998.715 / .750
Vertex CD-L2 ↓ (×10⁻⁴)10.5 / 10.418.9 / 20.266.1 / 62.5
Vertex CD-L1 ↓ (×10⁻⁴)16.4 / 16.127.6 / 29.596.5 / 92.0

Ablations trained to the full schedule (Table 4 metric) are shown below. Specifically, w/o vertex-link uses absolute-position RFF of the vertices in both the previous and the revised version.

VariantPRF1
Ours.9518.9902.9697
mean pooling.8866.9702.9185
w/o vertex-link.6584.8085.7179
w/o RoPE.4007.5138.4097

4. Test set (W2, Q7)

The 185 Objaverse test meshes form a held-out split, which we deduplicated against the training set with mesh statistics, render view image feature matching, VLM verification, and manual checks. The 500 Toys4K meshes, drawn at random among meshes with fewer than 20K vertices, are not part of our training data. Regarding the small matched subsets. They follow the published face caps of the autoregressive methods. This favors these methods. The supplementary also reports the comparison on the full sets, and the gaps widen there. Regarding the training splits of the baselines. Most of them are not released, so we cannot verify the overlap with our test meshes. Any such overlap would favor the baselines, not our method.

5. Vertex budget control (Q6)

The vertex budget is part of our problem setting and widely used in the industry, i.e., vertex budget is left for the user to define. We generated the 185 Objaverse meshes at vertex budget scales of 0.5, 0.75, 1.0, 1.25 and 1.5 under the new inference config.

Vertex budget scale0.50.751.01.251.5
CD-L2 (×10⁻⁴)68.864.559.761.362.8

The number of generated vertices rises monotonically with the scale, from 0.61 to 1.34 times the count at scale 1.0, and the surface stays close to the reference at every budget. A budget scaled up or down by 25% changes CD-L2 by 1.6 to 4.8 (×10⁻⁴), and the lowest error is reached at scale 1.0. A new figure in the revised PDF shows four Objaverse inputs generated at the five budgets.

6. Flow-based anchor generator (Q8)

From 15K to 115K training steps, the median CD-L2 (×10⁻⁴) of the flow-based anchor stayed between 76 and 78, so longer training does not close the gap, nor does inference config tuning improve it. Thus the stage 1 anchor generator remains the faster, lighter and better-performing choice.

7. Dual-grid processed shape input (Q9)

The frozen shape encoder of the flows reads points sampled from the original surface. This is standard practice in prior works, such as BPT (Weng et al., 2025), DeepMesh (Zhao et al., 2025), TreeMeshGPT (Lionar et al., 2025), FastMesh (Kim et al., 2026) and Nexus (Wang et al., 2026). We also ran the corresponding control experiment for MeshCarve, and the generation performance differs only marginally. Specifically, when it reads points from the dual-grid remesh instead, CD-L2 (×10⁻⁴) changes from 59.7 to 60.5 on Objaverse and from 54.6 to 54.2 on Toys4K.

Comments on the paper

1 comment

We thank the reviewers for their detailed reviews. We have optimized our inference configuration for faster generation and better generation quality without changing the methods and model weights reported in the submission. The changes are marked in blue in the revised PDF. We also thank the reviewers for pointing out that both the VAE and the scaled flow matching backbones can contribute to the performance improvements. We agree and would like to clarify that the VAE design is closely coupled with the flow matching backbone. Without the compression achieved by the VAE, training flow matching models can be quite costly and inference also takes longer. We have also included additional experiments sizing down the edge flow model that further ablate the VAE and flow model's contribution to the improvements. We also received feedback regarding LATO.2. We would like to point out that LATO.2 is a concurrent ICLR 2027 submission. Per the ICLR 2027 Reviewer Guidelines, our performance does not need to be compared against it. We nevertheless included it in our submission. We regard the performance differences to LATO.2 as a reference only. Though both are for artisan mesh generation, the focuses are different and the contributions also lie in different places.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.