Autoregressive models for 3D mesh generation suffer from a fundamental limitation: they flatten meshes into long vertex-coordinate sequences. This results in prohibitive computational costs, hindering the efficient synthesis of high-fidelity geometry. We argue this bottleneck stems from operating at the wrong semantic level. We introduce FACE, a novel Autoregressive Autoencoder (ARAE) framework that reconceptualizes the task by generating meshes at the face level. Our one-face-one-token strategy treats each triangle face, the fundamental building block of a mesh, as a single, unified token. This simple yet powerful design reduces the sequence length by a factor of nine, leading to an unprecedented compression ratio of 0.11, halving the previous state-of-the-art. This dramatic efficiency gain does not compromise quality; by pairing our face-level decoder with a powerful VecSet encoder, FACE achieves state-of-the-art reconstruction quality on standard benchmarks. The versatility of the learned latent space is further demonstrated by training a latent diffusion model that achieves high-fidelity, single-image-to-mesh generation. FACE provides a simple, scalable, and powerful paradigm that lowers the barrier to high-quality structured 3D content creation.
Figures & tables
Figure 2: The end-to-end pipeline of our FACE model. An encoder compresses the input point cloud into a latent VecSet. An autoregressive decoder then conditions on this VecSet, generating the mesh face-by-face. A Face Embedding layer encoder 9 tokens of each face and a CausalMLP head decodes each latent face token into 9 quantized coordinate tokens.
Method
MeshXL
MeshAnything
MeshGPT
PivotMesh
EdgeRunner
MeshAnything v2
DeepMesh
Nautilus
BPT
Mesh-Silksong
TreeMeshGPT
Ours
Compression Ratio ↓
1.00
1.00
0.67
0.67
0.47
0.46
0.28
0.27
0.26
0.22
0.22
0.11
Table 1: Comparison on mesh token efficiency. Our method achieves a state-of-the-art compression ratio, representing the mesh with the shortest sequence length.
Figure 3: Overview of our image-to-mesh generation pipeline. We first use the input image to condition a DiT model. The resulting latent VecSet is then fed into the Autoregressive Face Decoder to produce the final mesh.
Objaverse
Toys4K
Famous
Method
Hausdorff ↓
Chamfer ↓
Hausdorff ↓
Chamfer ↓
Hausdorff ↓
Chamfer ↓
MeshAnything
0.327
0.128
0.264
0.110
0.399
0.177
MeshAnythingV2
0.265
0.089
0.238
0.090
0.488
0.203
TreeMeshGPT
0.183
0.055
0.133
0.047
0.226
0.064
BPT
0.126
0.043
0.091
0.037
0.143
0.061
Ours (FACE)
0.090
0.041
0.067
0.033
0.077
0.049
Table 2: Quantitative comparison of mesh reconstruction quality on multiple datasets.
Figure 4: Qualitative comparison of mesh reconstruction quality on the Toys4K dataset.
Figure 5: Qualitative comparison of image-conditioned mesh generation.
Method (Order)
Hausdorff Distance ↓
Chamfer Distance ↓
BFS
0.728
0.528
DFS
0.171
0.077
ZYX-component
0.110
0.045
ZYX
0.103
0.047
Table 3: Ablation on mesh face ordering strategies.
Queries
Hausdorff Distance ↓
Chamfer Distance ↓
Learnable
0.132
0.058
Downsample
0.103
0.047
Table 4: Ablation on different queries of the shape encoder.
Method
Hausdorff Distance ↓
Chamfer Distance ↓
Parallel Decode
0.426
0.239
Attention-based
0.132
0.064
CausalMLP
0.103
0.047
Table 5: Ablation on coordinate decoding strategies.
Figure 6: Qualitative comparisons between the base and large model.