[CVPR 2025] FACE: Redefining Mesh Generation with the "One-Face-One-Token" Paradigm
FACE: A Face-based Autoregressive Representation for High-Fidelity and Efficient Mesh Generation
FACE is a novel Autoregressive Autoencoder (ARAE) for high-fidelity 3D mesh generation that introduces a "one-face-one-token" strategy. By representing each triangle face as a single unified token rather than individual vertex coordinates, it achieves a State-of-the-Art (SOTA) compression ratio of 0.11 and outperforms previous methods in reconstruction quality.
TL;DR
Current Autoregressive (AR) models for 3D meshes are "token-hungry," flattening geometry into massive vertex sequences that crush Transformer performance. FACE breaks this bottleneck with a deceptively simple insight: one face, one token. By elevating the semantic level of generation, FACE achieves a 0.11 compression ratio (2x better than the previous SOTA) and significantly higher reconstruction fidelity, paving the way for scalable, high-resolution 3D content creation.
The Bottleneck: Why Vertex Sequences Fail
Triangle meshes are the industry standard for 3D assets, but generating them directly is notoriously difficult. Existing AR models like MeshGPT or MeshXL treat meshes as a series of numbers (vertex coordinates).
The Math of Inefficiency: A mesh with faces typically results in a sequence length of (3 vertices per face 3 values per vertex). Since the computational cost of Transformers grows quadratically with sequence length (), doubling the mesh resolution increases the cost 16-fold. Most prior work tried to fix this with complex "traversal" rules to reuse vertices, but these are often brittle and break the global structure.
The FACE Insight: Elevating the Semantic Level
The authors of FACE argue that we are operating at the wrong semantic level. Instead of predicting coordinates, why not predict faces?
1. One-Face-One-Token
By projecting a 9D face vector into a single latent token, the sequence length drops by a factor of 9. This leads to a theoretical 81x reduction in self-attention cost.
2. The Architecture: ARAE
FACE uses an Autoregressive Autoencoder (ARAE) framework:
- Shape Encoder: Utilizes a VecSet architecture to compress point clouds into a global latent representation.
- Autoregressive Face Decoder: Generates tokens step-by-step.
- CausalMLP Head: A specialized head that decodes a single face token back into 9 coordinates. This is the "secret sauce"—it enforces internal dependency within a face, ensuring the three vertices actually form a coherent triangle.
Figure 1: The end-to-end FACE pipeline. A point cloud is encoded into a VecSet, guiding the face-by-face generation process.
Results: Efficiency Without Compromise
FACE doesn't just run faster; it looks better. By reducing the sequence length, the Transformer can focus its limited "attention budget" on higher-resolution global relationships rather than getting lost in a sea of coordinate tokens.
| Metric | MeshGPT | Mesh-Silksong | Ours (FACE) |
|---|---|---|---|
| Compression Ratio ↓ | 0.67 | 0.22 | 0.11 |
| Hausdorff Distance ↓ | 0.327 | 0.226* | 0.090 |
Note: Relative comparison highlights the massive jump in both efficiency and accuracy.
Qualititave Superiority
Compared to earlier SOTAs like MeshAnythingV2, FACE reconstructions are noticeably cleaner, capturing sharp edges and thin structures (like chair legs or tree branches) that previously resulted in holes or "melted" geometry.
Figure 2: Qualitative superiority of FACE in capturing intricate details on the Toys4K dataset.
Beyond Reconstruction: Image-to-Mesh
Because FACE learns a semantically rich latent space, it can be extended to generative tasks. By training a Latent Diffusion Model (DiT) on the VecSet tokens, the authors enable high-fidelity image-to-mesh generation.
Figure 3: Generating 3D meshes from single 2D images via Latent Diffusion in the FACE latent space.
Scaling Up
The paper concludes by scaling FACE to a 1.2 Billion parameter model. By increasing the quantization resolution to 1024, the "Large" model captures details indistinguishable from the ground truth, proving that the "Face-per-token" paradigm is the future of high-resolution 3D modeling.
Critical Analysis & Conclusion
Takeaway: FACE succeeds by simplifying the problem. By moving from vertex-level "pixels" to face-level "patches," it mimics the success of ViT in computer vision.
Limitations:
- Discrete Bounds: While 1024-resolution is high, it is still a voxel-like quantization. Truly continuous surfaces remain a challenge for quantized AR models.
- Sampling Density: The model relies on input point clouds; if a thin part (like a wire) isn't sampled well in the point cloud, the mesh output may struggle.
FACE is a milestone in 3D generative AI, proving that sometimes the best way to handle complexity is to move the goalposts to a higher level of abstraction.