Reinforcement learning (RL) has demonstrated remarkable success in text and image generation, yet its potential in 3D generation remains largely unexplored. Existing attempts typically rely on offline direct preference optimization (DPO) method, which suffers from low training efficiency and limited generalization. In this work, we aim to enhance both the training efficiency and generation quality of RL in 3D mesh generation. Specifically, (1) we design the first asynchronous online RL framework tailored for 3D mesh generation post-training efficiency improvement, which is 3.75× faster than synchronous RL. (2) We propose Advantage-guided Ranking Preference Optimization (ARPO), a novel RL algorithm that achieves a better trade-off between training efficiency and generalization than current RL algorithms designed for 3D mesh generation, such as DPO and group relative policy optimization (GRPO). (3) Based on asynchronous ARPO, we propose Mesh-Pro, which additionally introduces a novel diagonal-aware mixed triangular-quadrilateral tokenization for mesh representation and a ray-based reward for geometric integrity. Mesh-Pro achieves state-of-the-art performance on artistic and dense meshes.
Figures & tables
Figure 2: Architecture Overview. Mesh-Pro begins by sampling point clouds from the input dense and artist meshes. The features from the point cloud encoder are then passed to an auto-regressive Hourglass Transformer [ 18 ] for mesh decoding. This decoder is trained with truncation to output triangle-quad tokens. The pre-training objective is to reconstruct the input mesh. Subsequently, asynchronous ARPO is used for RL post-training to generate high-quality, well-structured meshes, guided by ray and topological rewards.
Figure 3: Diagonal-Aware Mesh Tokenization. “P” denotes vertex tokens. The minimum vertex always appears first in each face (i.e., lower coordinates). Triangles use padding tokens (“S”) at the end, while quads encode diagonal information in the fourth vertex via offset flag×2nbits ( flag∈{0,1,2} for Diagonals 1, 2, 3). The three edges of the first triangle (in green) can all potentially serve as diagonals. This defers the triangle-vs-quad decision to the last position, reducing prediction pressure.
Figure 4: Asynchronous Online RL Framework.
Figure 5: Qualitative comparison of Mesh-Pro with other methods.
Figure 6: Performance of asynchronous ARPO over training steps. Y-axes for “ ↓ ” metrics (BR, CD, HD) are intentionally inverted to align upward trends with performance gains (similarly below).
Figure 7: Performance curves of asynchronous RL methods and reward function over training steps.
Data Type
Dense Meshes
Artist Meshes
Metrics
CD ↓
HD ↓
BR ↓
QR ↑
US ↑
CD ↓
HD ↓
BR ↓
QR ↑
US ↑
MeshAnyv2 [ 8 ]
0.148
0.395
94%
54%
1.5
0.084
0.239
95%
57%
1.8
BPT [ 51 ]
0.109
0.284
72%
43%
2.5
0.046
0.131
50%
47%
2.5
DeepMesh [ 61 ]
0.351
0.635
91%
64%
2.8
0.336
0.584
64%
66%
2.7
FastMesh [ 22 ]
0.099
0.256
99%
4%
1.1
0.044
0.133
98%
18%
1.2
Mesh-RFT [ 33 ]
0.051
0.161
40%
65%
3.4
0.041
0.134
38%
69%
3.5
Table 1: Quantitative comparison on Dense and Artist Meshes.
Group
Advantage
Online
Pre-Start
CD ↓
HD ↓
BR ↓
QR ↑
US ↑
0.034
0.103
24%
55%
1.4
2
0.032
0.102
23%
70%
2.1
4
0.030
0.097
24%
73%
2.4
4
✓
0.029
0.093
23%
76%
2.9
4
✓
✓
0.033
0.103
23%
74%
2.7
4
✓
✓
✓
0.028
0.090
22%
81%
3.5
Table 2: Effectiveness analysis of each component of asynchronous ARPO. “*” denotes the pretrained model.
Figure 8: Visual comparison of different asynchronous online RL algorithms.
Metrics
CD ↓
HD ↓
BR ↓
QR ↑
US ↑
Pretrained Model
0.034
0.103
24%
55%
0.4
with Async DPO
0.028
0.093
23%
75%
1.9
with Async GRPO
0.033
0.096
23%
68%
1.3
with Async ARPO
0.028
0.090
22%
81%
2.4
Table 3: Quantitative comparison of ARPO with DPO and GRPO.
Data Scale
BR ↓
US ↑
400
22.35%
0.86
700
22.07%
1.01
1000
21.83%
1.13
Table 4: Performance with scaled data in asynchronous ARPO.
Tokenizer
Async ARPO
CD ↓
HD ↓
BR ↓
QR ↑
US ↑
Previous
w/o
0.056
0.122
38%
72%
1.2
with
0.071
0.160
46%
80%
1.8
Ours
w/o
0.034
0.103
24%
55%
0.9
with
0.028
0.090
22%
81%
2.1
Table 5: Comparison of different mesh tokenizations.
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 9: Python Code of ARPO.
Figure 10: Illustration of ray casting integrity check. When a ray is cast toward the mesh from a given direction, it may pass through the broken region surrounding a “bad face”, leading to a direct back-face hit on the mesh. The number of bad faces Nbf reflects the extent of geometric breakage in the generated mesh.
Figure 11: Boundary edge–based reward leads to misjudgments of the multi-component object. When a mesh consists of multiple components, boundary edges (highlighted by the green line) often appear between components. However, this mesh is still a good output and should be encouraged. In Mesh-Pro, the ray-based reward does not suffer from this issue.
Figure 12: Illustration of quad rings and quad lines.
Figure 13: Distribution of face count (consisting of a mixture of triangles and quadrilaterals) in Mesh-Pro predictions. Point clouds are sampled from dense meshes and artist meshes. The average face count is approximately 8k.
Figure 14: Training loss curves for DPO, GRPO, and ARPO.
θsucc
0.01
0.05
0.1
MeshAnyv2
94%
78%
62%
DeepMesh
91%
81%
69%
FastMesh
99%
98%
95%
Mesh-Pro
22%
7%
1%
Appendix
Table 6: Broken ratio sensitivity analysis.
Configuration
Baseline
+KV caching
+ CUDA graph
Token/s ↑
40
110
310
Appendix
Table 7: Runtime Analysis of Mesh-Pro.
Data Type
Dense Meshes
Artist Meshes
Metrics
CD ↓
NC ↑
F1 ↑
CD ↓
NC ↑
F1 ↑
MeshAnyv2 [ 8 ]
0.148
0.679
0.064
0.084
0.785
0.134
BPT [ 51 ]
0.109
0.737
0.151
0.046
0.876
0.311
DeepMesh [ 61 ]
0.351
0.535
0.027
0.336
0.579
0.082
FastMesh [ 22 ]
0.099
0.714
0.112
0.044
0.849
0.254
Mesh-RFT [ 33 ]
0.051
0.809
0.208
0.041
0.885
0.328
Appendix
Table 8: More quantitative comparison on Dense and Artist Meshes.
Figure 15: Multi-view renderings of meshes generated by Mesh-Pro. Mesh-Pro produces diverse meshes with robust geometric integrity and high topological quality.
Figure 16: Qualitative comparison with the closed-source commercial quad mesh generation method Tripo and Hunyuan3D. Mesh-Pro exhibits higher geometric consistency, richer details, and superior topology and edge-flow quality.
Figure 17: High-quality native quad-dominant topology generated by Mesh-Pro demonstrates robust performance in downstream tasks, such as UV unwrapping, texture painting, and animation.