acceptodds
Under review as a conference paper at ICLR 2027

Taming TRELLIS.2: Topology-Aware Online Post-Training for 3D Mesh Generation

Abstract

Directly generating 3D assets is essential for scalable 3D content creation. Recent latent flow-based 3D generative models, such as TRELLIS.2, achieve impressive generation quality and efficiency by employing expressive 3D representations that support arbitrary topology. However, such topological flexibility comes at the cost of losing guarantees on topological regularity, potentially introducing unintended non-manifold structures when regular topology is desired. We propose TopoFT, a topology-aware fine-tuning method for online post-training of pretrained 3D generators that improves topological regularity without modifying the underlying O-Voxel representation or restricting its support for arbitrary topology. Our diagnostic analysis identifies Stage-2 shape-SLat generation as a major source of topology irregularity. Since topology is evaluated globally while individual irregularities are spatially localized, TopoFT decomposes each mesh-level group-relative advantage into additive token-wise attributions that exactly preserve the global advantage. An anchored dual-branch flow objective then reinforces favorable local trajectories while stably suppressing unfavorable ones. Across four unseen benchmarks, TopoFT consistently reduces non-manifold edges, non-manifold vertices, and boundary edges without systematic degradation in geometric fidelity. These improvements translate into substantial downstream benefits: the target-reaching rate under aggressive QEM simplification increases from 52.7–91.0% to 95.8–99.4%, while the open-wall path ratio in downstream slicing decreases from 15.7–22.9% to 8.0–14.5%.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Generation, Image-to-3D Generation, Mesh Generation, Flow Matching, Mesh Topology

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_qVbiOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduce a method for improving the topology of meshes generated by Trellis2. It keeps the original OVoxel representation and the decoder, states that Stage-2 shape-SLat generation is the major source of topology irregularity. The method decomposes each mesh-level group-relative advantage into additive token-wise attributions that exactly preserve the global advantage, then use an anchored dual-branch optimization to further improve the result quality. The methtod is evaluated on many benchmarkes and have reduced in non-manifold and boundary elements., while still maintain good surface distance metric.

Critical strengths

The method is well motivated. With the Trellis2 representation, the main issue is the open-edges and topology issues including holes in the generated mesh. The diagnostic is reasonable. The solution is also follows the supported finding in the dignostic process and the method is reasonable.

The evluation result is promising. The result show consisten imporvements over the originial trellis 2 model, while the quality is largely remain keept.

Critical weaknesses

The paper is awkwardly written. It more liked a generarted report than a research paper. Some terms like 'group-relative advantage' and 'token-wise advantage attributions' reads strange to me, maybe I lack some domain konwledge here.

The overall contribution is somewhat limited and a bit ad-hoc. There should be more systematic way to solve the problem, the currrent solution is more like analysis the Trellis2 and fixed the problem as what is shows.

The visual evaluaton is only on simple shapes and low token counts. I'd like to see stress test on high token counts and complex shapes.

Questions for authors

My first question is how much AI is being used to generate the submission.

Can the author discuss why the related work is put at the last? and the intention of having so many packed math in sec 3.3?

Can the author speaks more on how he evaluaton is conducted? is this performed on generation taks (from rendered image?) or on reconstruction tasks?

4 comments

We thank the reviewer for their thoughtful comments and valuable questions.

W1. Writing and Terminology

A. Our paper progresses from controlled diagnosis (Sec. 3.2) to method development (Sec. 4) and experimental validation (Sec. 5).

Regarding terminology, group-relative advantage follows the established GRPO formulation introduced in DeepSeekMath (Shao et al., 2024) and subsequently adopted for flow-matching models in Flow-GRPO (Liu et al., NeurIPS 2025). It measures each sample's relative quality within a group.

Token-wise advantage attribution decomposes the sample-level advantage into token contributions whose sum equals the original advantage, as defined in Eq. (10). We use attribution to emphasize this additive relationship.

References

[1] Shao et al., DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models, 2024.

[2] Liu et al., Flow-GRPO: Training Flow Matching Models via Online RL, NeurIPS 2025.

The terminology definitions are provided in the Method section.

W2. Contribution and Generality

A. Recent diffusion- and flow-matching-based 3D generation frameworks have made substantial progress in generation quality and efficiency. However, the generated meshes can still contain topology irregularities. We address this common problem through post-training, with two main components.

  1. Mesh-to-token topology attribution. Applying the same mesh-level feedback to all tokens does not distinguish which regions contribute to topology irregularities. We attribute mesh-level topology irregularities to individual latent tokens, providing localized optimization signals. The same attribution principle could apply beyond TRELLIS.2 to other 3D generators whose decoded mesh regions can be linked to latent tokens.

  2. Anchored dual-branch optimization. The positive branch reinforces token targets with favorable topology feedback. The negative branch uses the current model prediction as a reference to discourage fitting targets with unfavorable feedback. We incorporate positive and negative topology feedback into flow matching through an objective that remains bounded below. The optimization uses flow predictions and signed feedback, so it could also be applied to latent flow-matching models beyond TRELLIS.2.

We use TRELLIS.2, a representative architecture, to evaluate the proposed approach. TopoFT improves topology across four test benchmarks while maintaining comparable geometric fidelity (Tables 1–2). Although our experiments focus on TRELLIS.2, the two components are not inherently restricted to this architecture and could be adapted to other latent flow-based mesh generators.

W3. High-Token-Count and Complex-Shape Evaluation

A. We have added a supplementary comparison video featuring more challenging geometries and higher token counts, such as engines and castles (see the newly attached video).

It reports normalized rates and relative reductions for non-manifold edges, non-manifold vertices, and boundary edges. The results show that TopoFT substantially reduces topology irregularities even for complex shapes, while preserving geometric fidelity.

Q1. Use of AI

A. As stated in the AI Use Statement of our paper, we used ChatGPT as an assistive tool, primarily for language editing and literature searches.

We made all methodological decisions, established the experimental protocols, and interpreted the results. We obtained all reported results using our own experimental pipelines.

Q2. Paper Organization and Mathematical Presentation

A. We originally placed Related Work after the experiments to highlight our contributions earlier. Relevant prior work is also discussed in the Introduction. This structure is also used in recent papers such as SAM 3D (CVPR 2026) [1], FlowR (ICCV 2025) [2], and two ICLR 2025 Outstanding Papers [3, 4]. We have adopted the suggestion and moved the Related Work section after the Introduction in the revised manuscript.

The mathematics in Sec. 4.3 explains why we use the anchored dual-branch objective. Direct negative weighting makes the flow-matching loss unbounded below. Our anchored negative branch keeps the loss bounded below while preserving the signed gradient at the sampling-model prediction. The proposition proves this property.

References

[1] Chen et al., SAM 3D: 3Dfy Anything in Images, CVPR 2026.

[2] Fischer et al., FlowR: Flowing from Sparse to Dense 3D Reconstructions, ICCV 2025.

[3] Fang et al., AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models, ICLR 2025 Outstanding Paper.

[4] Qi et al., Safety Alignment Should Be Made More Than Just a Few Tokens Deep, ICLR 2025 Outstanding Paper.

Q3. Evaluation Setup and Experimental Design

A. Our main experiments evaluate single-image-conditioned 3D generation on four test benchmarks. Each input image is rendered from a ground-truth mesh, which is never provided to the generator and is used only for rendering and geometric-fidelity evaluation.

The reconstruction pipelines in Sec. 3.2 are separate diagnostic experiments. The test benchmarks, evaluation metrics, and protocols are detailed in Sec. 5.1 and Appendix B.5.

The evaluation setup is described in the "Experimental Setup" subsection.

Thanks for the updates. The visual result in the video although provide better numerics, the visual resutls are not nesseary improving from TRELLIS.2. There are still visible holes in the TopoFT, espeically the last three examples in the video, it is hard to say the TopoFT results are improved from TRELLIS.2.

This is a good submission, but still not conviencing enough. Especially the cliam in the rebuttal the method works beyond fixing TRELLIS.2 is a bit bold.

Thank you for your further comments.

1. Visual Result

To make the visual differences easier to examine, we have added enlarged comparisons of all examples from the supplementary video (Figs. 8–11 in the Appendix), including the last three examples mentioned in your comment. We did not select examples based on the extent of improvement.

The enlarged views show that TopoFT reduces many small holes, fragmented surfaces, and local irregularities present in TRELLIS.2, including on complex structures.

We acknowledge that some irregularities remain, particularly in highly complex shapes. TopoFT aims to reduce topology irregularities rather than guarantee fully manifold or watertight meshes.

2. Generality beyond TRELLIS.2.

To further investigate the applicability of TopoFT beyond TRELLIS.2, we conducted an additional experiment on LATO.2 [1], a flow-based 3D mesh generator developed as an extension of LATO (ICML 2026) [2].

Unlike TRELLIS.2, which generates shape-SLat features and decodes them into meshes through O-Voxel, LATO.2 separates vertex generation from topology generation. Its V-Flow generates vertex positions, while T-Flow generates per-vertex topology latents that are decoded into mesh connectivity.

We adapted both two components of TopoFT to LATO.2. Specifically, we attribute mesh topology irregularities to the corresponding T-Flow tokens and apply anchored dual-branch optimization to its velocity predictions. We keep V-Flow frozen and fine-tune only T-Flow.

Due to the limited rebuttal time, we performed only 400 optimization steps using the Objaverse-XL training pool, which is separate from the GSO test set. We then evaluated pretrained LATO.2 and its TopoFT variant on GSO.

Topology MetricPretrained LATO.2TopoFT
NME (%) ↓13.095.47
NMV (%) ↓1.210.26
BND (%) ↓4.421.43
Geometric FidelityPretrained LATO.2TopoFT
Chamfer (×10⁻²) ↓0.610.60
P95 (×10⁻²) ↓1.201.15
F-score @ 0.01 (%) ↑89.5990.39

Despite only 400 optimization steps, TopoFT substantially reduces all three topology irregularity ratios while maintaining comparable geometric fidelity. More details are in Appendix F.

These preliminary results provide evidence that both components of TopoFT can be transferred to a different mesh-generation architecture. While broader validation is still needed, they suggest that our approach has the potential to extend to other flow-based 3D mesh generators.

We hope this additional experiment helps clarify the scope and potential generality of our method.

[1] Long, H., et al. (2026). LATO.2: Factorized 3D Mesh Generation with Vertex and Topology Flow. arXiv:2607.10623.

[2] Zhao, T., et al. (2026). LATO: 3D Mesh Flow Matching with Structured TOpology Preserving LAtents. Proceedings of the 43rd ICML.

Reviewer_Z7JaOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TopoFT post-trains TRELLIS.2 using localized topology feedback and an anchored flow-matching objective to reduce mesh irregularities while preserving geometry.

Critical strengths
  1. There are consistent topology improvements across four benchmarks and three training seeds.
  2. Token-level attribution outperforms mesh-level feedback, and the simplification and slicing results demonstrate practical benefits.
  3. Complexity controls help rule out trivial gains from reducing mesh size.
Critical weaknesses
  1. QEM success rates exclude failed runs, the overall reliability is unclear.
  2. The anchored objective is also not independently ablated.
Questions for authors
  1. Are predictions and ground truth in the same orientation, and what explains the low F-scores?
  2. What are the QEM failure rates and target-reaching rates over all attempted meshes?
  3. Can you isolate the benefit of the anchored objective while keeping online sampling and token attribution fixed?
3 comments

We thank the reviewer for their thoughtful comments and valuable questions.

Q1. Orientation Alignment and F-scores.

A. Thanks for this constructive comment. Raw predictions and ground-truth meshes are not guaranteed to have the same orientation. These orientation differences explain the relatively low F-scores in our initial evaluation.

We re-evaluated F-score, Chamfer distance, and P95 distance after rigidly aligning predictions with ground truth.

DatasetMethodChamfer (×10⁻²) ↓P95 (×10⁻²) ↓F-score at 0.05 (%) ↑
GSOTRELLIS.23.5212.6180.51
TopoFT3.4912.5280.86
Fusion360TRELLIS.24.1015.4077.11
TopoFT4.0315.1277.52
OmniObject3DTRELLIS.22.327.6989.89
TopoFT2.317.6389.93
Toys4KTRELLIS.23.2511.8582.71
TopoFT3.2411.7382.88

Under the revised evaluation protocol, the overall trend remains unchanged: TopoFT maintains similar geometric fidelity to pretrained TRELLIS.2 across all four datasets.

The updated results are reported in Tables 2 and 8, with the normalization and registration details in Appendix B.5.1.

Q2. QEM Reliability over All Attempted Meshes.

A. The pretrained TRELLIS.2 and Full TopoFT have the same failure rate of 0.08%. All failures arise from the shared Stage-1 model generating an empty set of active voxels. In these cases, no mesh can be generated, so Stage-2 generation and subsequent QEM simplification are skipped. These failures occur before our fine-tuned Stage-2 model is used and are unrelated to our method.

DatasetMethodTarget reached (%) ↑
GSOTRELLIS.273.90
TopoFT98.83
Fusion360TRELLIS.290.81
TopoFT99.25
OmniObject3DTRELLIS.252.66
TopoFT99.30
Toys4KTRELLIS.262.09
TopoFT95.75

We evaluate QEM reliability over all samples, counting missing or invalid outputs as failures. TopoFT achieves higher target-reaching rates across all four datasets, showing improved QEM simplification reliability.

The results are reported in Appendix A.7, and the evaluation protocol for Tables 3 and 9 is clarified in Appendix B.5.2.

Q3. Independent Ablation of the Anchored Objective.

A. Please see our General Response for the controlled branch ablation results.

Thanks for the response and I will keep my rating

Thank you for taking the time to review our work and consider our response. We appreciate your positive assessment of the consistent topology improvements, the benefits of token-level attribution, and the effectiveness of our complexity controls. We are also glad that you recognize the practical benefits demonstrated by our simplification and slicing experiments.

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

This paper proposes TopoFT, an online post-training method for improving the topology of meshes generated by TRELLIS.2. A stage-wise analysis identifies Stage-2 as a major source of irregularities. TopoFT assigns group-relative mesh feedback to spatial tokens and updates the model with an anchored dual-branch flow loss. Experiments report improved topology across datasets, alongside gains in mesh simplification and slicing.

Critical strengths
  1. Well-motivated design. The stage-wise diagnosis supports focusing on Stage-2. Evaluation across four datasets, mesh simplification, and slicing shows benefits beyond a single topology metric.
  2. Clear optimization idea. The token attributions sum to the global advantage, while the anchored negative branch avoids the unbounded objective that would result from directly applying negative weights to squared flow-matching errors.
Critical weaknesses
  1. Efficiency and mechanism evidence is incomplete. The paper does not report training time or the variation among eight candidates for the same condition, and it does not isolate the contribution of the anchored negative branch.
Questions for authors
  1. How much do the eight candidates sampled under the same condition differ in topology cost? Could you report the distribution of within-group standard deviations, the fraction of near-zero-variance groups, and the corresponding attribution magnitudes?
  2. Since v_old is a prediction from the frozen sampling model rather than a quality-selected target, could you clarify when the negative branch helps and whether it can disrupt good generation modes?
  3. Could you compare the positive branch alone, the anchored negative branch, and the full method under the same online sampling budget? This would help separate their contributions, since the DSO comparison changes several factors at once.
  4. What are the wall-clock time and GPU-hours for a training run? A breakdown of Stage-2 sampling, mesh decoding, topology evaluation, and optimization would help assess the cost of online feedback.
3 comments

We thank the reviewer for their thoughtful comments and valuable questions.

Q1. Within-Group Cost Variation and Token Attribution

A. The table shows the distribution of within-group topology cost standard deviations (STDs), with the quartiles presented in ascending order.

Statistic25th percentileMedian75th percentile
Cost STD7.69×10−47.69\times10^{-4}1.57×10−31.57\times10^{-3}2.86×10−32.86\times10^{-3}

CV=STDMean,median⁡(CV)≈1.38×10−1.\displaystyle \mathrm{CV}=\frac{\mathrm{STD}}{\mathrm{Mean}},\qquad \operatorname{median}(\mathrm{CV})\approx1.38\times10^{-1}.

The median coefficient of variation (CV) is (1.38×10−11.38\times10^{-1}), where CV is calculated as each group's STD divided by its mean cost. These results show that candidates sampled under the same condition generally differ in topology cost.

We define near-zero variation as an STD below 0.01×0.01\times the median STD. Only 0.0381% of groups meet this criterion. The near-zero-variance groups have a mean absolute token attribution of 4.28×10−34.28\times10^{-3}, compared with 4.88×10−34.88\times10^{-3} for other groups. Thus, near-zero-variance groups are rare, and their attribution magnitudes are similar to those of other groups.

The within-group statistics are reported in Appendix A.3, and the SD distribution is shown in Appendix Figure 5.

Q2. Effectiveness of the Negative Branch

A. The negative branch provides an explicit training signal against fitting sampled targets with unfavorable token-level topology feedback. The sampling prediction (voldv_{\mathrm{old}}) serves as the reference for the current model’s behavior. The residual (v⋆−voldv^\star-v_{\mathrm{old}}) represents the correction that would move this prediction toward the sampled flow-matching target. For a token with negative attribution, we reverse this correction and construct

+ a\left(v^\star-v_{\mathrm{old}}\right), \qquad a<0.$$ As illustrated in **Fig. 3** (the revised manuscript), **the resulting target lies in the opposite direction from ($v_{\mathrm{old}}$)**, discouraging the model from matching the sampled target at that token. The anchor defines the reference for this correction; the topology feedback determines which sampled targets should be reinforced or discouraged. The controlled ablation in **Table 1** (the revised manuscript) demonstrates the practical effect: **the negative branch reduces all three topology irregularity ratios across the four datasets**, and **combining it with the positive branch further improves the results**. To assess potential side effects on geometry, Table 2 and Appendix A.5 compare the full model with pretrained TRELLIS.2 and show comparable geometric fidelity across the four benchmarks. **Q3. Controlled Ablation** **A.** Please see our General Response for the controlled branch ablation results. **Q4. Training Runtime and Computational Cost** **A.** A complete training run takes **22.97** hours on two NVIDIA RTX PRO 6000 Blackwell GPUs, corresponding to **45.94** GPU-hours. Topology evaluation uses 24 CPU workers (2 × Intel Xeon Platinum 8470Q, 12 workers per GPU rank). The wall-clock breakdown is shown below. | Component | Total (h) | Share (%) | | --- | ---: | ---: | | Stage-2 sampling | 13.13 | 57.2 | | Mesh decoding | 0.60 | 2.6 | | Topology evaluation (exposed wait) | 1.47 | 6.4 | | Optimization | 3.44 | 15.0 | | Validation | 1.32 | 5.8 | | Other overhead | 3.00 | 13.1 | | Total | 22.97 | 100 | **Cost of Online Feedback.** The dominant cost comes from Stage-2 sampling (57.2%). Topology evaluation (TE) runs on CPU workers in parallel with GPU computation, but evaluations that outlast GPU computation cause additional waiting time. **Mesh decoding (2.6%) and TE wait (6.4%) together account for only 9.0% of total training time.** The total runtime is reported in Sec. 5.1, with the breakdown in Appendix B.3.

My concerns have been addressed and I will maintain my rating.

Thank you for taking the time to review our work and consider our response. We appreciate your positive assessment of our stage-wise analysis and optimization approach, and we’re glad our response addressed your concerns.

Comments on the paper

2 comments

We thank all reviewers for their careful reading and constructive feedback. We have revised the paper, responded to each reviewer, and added an ablation study of the positive and negative branches in Sec. 5.2 and Table 1.

We compared the Positive Branch, Negative Branch, and Full Method using the same training settings, differing only in the enabled loss branches.

DatasetMethodNME (‰) ↓NMV (‰) ↓BND (‰) ↓
GSOTRELLIS.213.200.8810.23
Positive Branch12.390.8710.24
Negative Branch8.590.447.81
Full Method3.590.105.23
Fusion360TRELLIS.27.460.214.99
Positive Branch6.730.204.94
Negative Branch4.420.123.95
Full Method2.230.042.75
OmniObject3DTRELLIS.220.961.3014.89
Positive Branch19.521.3515.24
Negative Branch10.640.459.33
Full Method3.810.064.77
Toys4KTRELLIS.220.651.2612.29
Positive Branch19.381.2912.46
Negative Branch13.040.8310.43
Full Method5.440.236.77

The Positive Branch performs similarly to TRELLIS.2 because positive advantages are computed relative to samples from the current model and do not guarantee good absolute topology quality. As a result, it can reinforce poor samples without suppressing poor generation behaviors.

The Negative Branch reduces all three topology irregularity ratios across the four datasets. The Full Method achieves the best results, showing the importance of negative feedback and the benefit of combining both branches.

Hi reviewers #Z7Ja and #qVbi, 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

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.