Mesh-RFT: Enhancing Mesh Generation via Fine-grained Reinforcement Fine-Tuning

We thank you for your strong support and insightful feedback. We address the weaknesses you raised below.

1. On the Justification of Topology Score (TS) Weights

Response: Our TS weights were optimized through iterative industrial benchmarks. To provide empirical support, our ablation study on 100 Hunyuan2.5 meshes shows that weight adjustments primarily influence local topological characteristics, not global geometry (HD remains stable, with Δ < 0.002).

This analysis reveals three key findings:

1. Weight sensitivity is local and topological: Adjustments alter face-type distribution without significantly affecting the global shape.
2. Equal weights impair quad merging: This configuration often results in non-coplanar triangle pairs, hindering quad conversion.
3. Higher w₂ improves aesthetics and utility: It boosts user preference by producing more rectangular quads, which streamline UV unwrapping workflows.

Thank you for highlighting this nuance. Based on these findings, we will adopt the High w₂ configuration in our final model to optimize for both user preference and practical utility.

2. On the Generality of Quad-Based Metrics

Response: We agree our metric is quad-biased, and this is a deliberate design choice to align with industry standards in pipelines like gaming and animation, where quad-dominant topology is critical for stable deformations during rigging, efficient UV unwrapping, and artist-friendly edge-loop editing. Our focus on generating "production-ready" assets therefore necessitates this optimization for high-quality quad topology.

3. On Computational Cost

Response: We agree the pre-training cost is substantial, and we made this investment to critically improve model stability and reduce generation failures. Our experiments show a direct correlation between extended pre-training and a lower "breakage rate"—the percentage of meshes with critical artifacts.

*Breakage Rate defined as the percentage of meshes with non-manifold edges and holes on the test set.

Crucially, the M-DPO fine-tuning—which contains the core contributions of our paper—adds only 8 hours of GPU time, making the refinement stage highly accessible. We will release the robust 10-day pre-trained model to lower the barrier to entry for the research community.

We appreciate your insightful feedback. We address your concerns below.

1. On Evaluation Bias and Cherry-Picking

Response: We appreciate your concern. A primary goal of our extensive pre-training was to enhance model stability and success rate, allowing us to consistently generate high-quality 10k-face meshes in a single pass without needing to cherry-pick from multiple runs. Our RL phase further refines topological quality to facilitate clean quad conversion. To demonstrate robust generalization, we evaluated our model on the complex examples from the first page of the TRELLIS website. While baseline models struggled to handle these cases, ours produced reasonable results, even for high-complexity meshes requiring up to 20k faces. Though our topology quality decreases at such high face counts, it remains significantly superior to other methods, as shown below.

2. On Compatibility with Special Topologies

Response: This is an excellent technical point. Our quad-biased Topology Score (TS) is a deliberate design choice to align with production standards in game design, where quad-dominant meshes are heavily favored. Our M-DPO framework is robust to this, as it balances both global and local quality scores, preventing minor topological constraints from degrading the overall quality of a complex mesh. For instance, in the car wheel example (Figure 11), while the central hub is formed with only two triangles, the model correctly generates a clean, circular structure for the outer tire. We have also tested this on other assets like headphones and observed that circular structures are often composed of triangles arranged in a way that facilitates easy conversion into clean rectangular patches. This results in a topology that is highly acceptable for artists and requires only minimal edits to finalize.

3. On Integrating M-DPO with Compact Tokenizers

Response: This is a key question regarding the trade-off between efficiency and generative performance. Our choice of an uncompressed representation is deliberate, as the hierarchical nature of our Hourglass Transformer naturally preserves geometric semantics from the coordinate-level up to the face-level. In contrast, compact methods like BPT, while enabling faster convergence by fitting more faces into the context window, often disrupt this inherent geometric structure. This creates a clear trade-off: our approach requires more extensive pre-training but achieves superior final performance by maintaining high-fidelity representations. The table below quantifies this relationship.

For researchers wishing to apply our Mesh-RFT framework to compressed formats, we suggest elevating the masking unit from the face-level to the patch-level. However, this may dilute the fine-grained refinement, as a single patch could contain a mix of both high- and low-quality faces, likely making the optimization less precise than our face-level approach. Nevertheless, it represents a promising direction for future research, which we will detail in our final paper.

We appreciate your insightful feedback. We address your concerns below.

1. On the Justification of Topology Score (TS) Weights

Response: Our TS weights were optimized through iterative industrial benchmarks. To provide empirical support, our ablation study on 100 Hunyuan2.5 meshes shows that weight adjustments primarily influence local topological characteristics, not global geometry (HD remains stable, with Δ < 0.002).

This analysis reveals three key findings:

1. Weight sensitivity is local and topological: Adjustments alter face-type distribution without significantly affecting the global shape.
2. Equal weights impair quad merging: This configuration often results in non-coplanar triangle pairs, hindering quad conversion.
3. Higher w₂ improves aesthetics and utility: It boosts user preference by producing more rectangular quads, which streamline UV unwrapping workflows.

Thank you for highlighting this nuance. Based on these findings, we will adopt the High w₂ configuration in our final model to optimize for both user preference and practical utility.

2. On the In-Depth Analysis of the Local Mask

Response: We appreciate the suggestion for a deeper analysis of our M-DPO mechanism. We address your specific concerns below:

1. On Degrading High-Quality Regions: Our M-DPO framework is explicitly designed to prevent this degradation by selectively applying the preference loss. As defined in Equation 3, the optimization has two distinct targets: it reinforces the high-quality regions of the preferred mesh (by applying the mask φ(M+)) while simultaneously penalizing the low-quality regions of the rejected mesh (by applying the inverted mask 1−φ(M-)). The flawed parts of the winner and the high-quality parts of the loser are both excluded from the loss calculation. Therefore, existing high-quality structures are not only preserved but are actively rewarded, ensuring that fine-tuning consistently leads to a net improvement in mesh quality.
2. On Overfitting to the Masking Strategy: Our extensive experiments show no evidence that the model overfits to the specific mask generation strategy. Because the model is trained to recognize and generate good topology (via the positive sample term) while unlearning bad topology (via the negative sample term), it learns the underlying principles of topological quality. This generalizes well to unseen mesh configurations, and the global structure is preserved because the fine-tuning process is guided by a robust pre-trained reference model.
3. On Soft vs. Hard Masks: We investigated the efficacy of soft masks versus our binary (hard) approach. For this experiment, we generated soft masks by normalizing each face's topology score into a continuous weight (e.g., via a sigmoid function centered at our quality threshold of 70). As shown below, the hard mask provides superior performance and faster convergence. We hypothesize this is because a binary mask offers a clear, unambiguous signal to the model—"reinforce this good region" or "penalize this bad region"—whereas a soft mask can introduce noisy gradients from moderately-scored regions, slowing down the learning process.

3. On Training Stability and Convergence

Response: We thank you for this valid concern. After initial hyperparameter exploration, we have achieved a stable and reproducible training configuration. The key adjustments were:

1. reducing the learning rate to 5e-7 (significantly lower than the pre-training rate of 1e-4),
2. increasing the DPO beta to 0.5 to stay close to the robust reference model,
3. using a dataset of 10k preference pairs, which we found sufficient for convergence without overfitting, and applying a label smoothing rate of 0.3 to the DPO loss.

With these settings (on 64 GPUs, total batch size 64), the training loss remains stable at ~0.693 for the first 2k steps, then steadily decreases to ~0.6 by the 4k step mark. Concurrently, the reward margin (reward_chosen - reward_rejected) begins to climb from 0 to ~0.7 after 2k steps, indicating successful and stable policy optimization.

We appreciate your insightful feedback. We address your concerns below.

1. On the Fairness of Comparison Against DeepMesh

Response: We agree the most critical comparison is between our proposed M-DPO and a standard global DPO when applied to the same strong backbone. As shown in our paper's Table 2, our method already outperforms a rule-based global DPO (S-DPO). To further demonstrate this, we conducted an additional ablation study including a global DPO trained on human preference pairs (H-DPO). The results below show that our M-DPO is superior to both global DPO variants. Interestingly, the rule-based S-DPO slightly outperforms the H-DPO, likely because our automated scoring provides more consistent preference signals than subjective human ratings.

2. On Lacking Qualitative Analysis for Metric Contributions

Response: While we cannot add new figures during the rebuttal, the quantitative ablation study below demonstrates the contribution of each component. As shown, DPO effectively learns the specified preference rules. Introducing the HD metric improves geometric consistency. Further incorporating our topological scores enhances mesh integrity by reducing holes and improving quad-convertibility. Finally, our full M-DPO method leverages local masking to refine fine-grained details, achieving the best overall performance. For a more intuitive understanding of how different topological scores impact mesh quality, please refer to our response to Reviewer dibn, Question 1, where we provide a detailed breakdown.

3. On Lacking Examples for Downstream Applications

Response: We agree that illustrating downstream impact is crucial. While a full pipeline is beyond this paper's scope, clean, quad-dominant topology is the industry standard for 3D artists because it saves significant manual effort and is essential for stable rigging and skinning, preventing artifacts like "skin tearing" during animation. Our method generates assets that are closer to this production-ready standard, directly addressing issues seen in baselines (e.g., the damaged foot in DeepMesh's result in Figure 5, row 2). To make this tangible, we commit to adding a dedicated section in the appendix of our final version that qualitatively demonstrates these benefits by showing a side-by-side skeletal deformation comparison between a mesh with good topology and one with poor topology.

4. On the Validity of the User Study

Response: Mesh completeness is a critical aspect of quality, which is why we dedicated substantial computational resources to pre-training to ensure model stability. However, to isolate the perceptual benefits of our topological improvements alone, we conducted a targeted user study where participants were shown only cropped local regions from our results vs. DeepMesh and asked to rate the topology. The results confirm a strong user preference for our method's fine-grained quality, proving the gain is not just from better completeness.