Table 2: Comparison with audio-driven facial animation methods. The best LVE value is marked in bold. “ z FACS GT +NFR” indicates the upper bound of the performance of our method.
| Method | Mesh-agnostic | LVE ↓ (×10−3) |
| CodeTalker | ✗ | 1.5927 ± 0.8608 |
| Faceformer | ✗ | 1.4854 ± 0.9858 |
| Ours | ✓ | 1.1776 ± 0.7796 |
| z FACS GT + NFR | - | 1.0642 ± 0.7425 |