Research map / Video / Video generation
UniTalker: Scaling up Audio-Driven 3D Facial Animation through A Unified Model
Xiangyu Fan, Jiaqi Li, Zhiqian Lin, Weiye Xiao, Lei Yang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train one UniTalker on all of A2F-Bench instead of BIWI only (ablation) | UniTalker-B · BIWI-Test-A | LVE (x10^-4): 4.3681 -> 4.0804 | Improved |
| scale training to A2F-Bench for small datasets (ablation) | UniTalker-L · BIWI and Vocaset | LVE: 4.279 -> 3.859 (BIWI) | Improved |
| mix large HDTF with other datasets (ablation) | UniTalker-L · D3 (3D-ETF-HDTF) | LVE (x10^-6 m^2): 8.445 -> 8.991 | Hurt |
| mix Chinese speech with other datasets (ablation) | UniTalker-L · D6 (Chinese speech) | LVE (x10^-5 m^2): 1.043 -> 1.145 | Hurt |
| fine-tune the pre-trained UniTalker on each seen dataset | UniTalker-L · A2F-Bench (8 datasets) | LVE: average LVE drop 6.3% | Improved |
| initialize from pre-trained UniTalker encoder instead of Wav2vec2-xlsr-53 for unseen annotation | BIWI fine-tuning · BIWI (D0) subsets | LVE (x10^-4): 4.249 -> 4.197 with half the data | Improved |
| multi-head model instead of preprocessing all annotations to FLAME vertices (ablation) | UniTalker · VOCA-Test (D1) | LVE (x10^-6 m^2): 8.7532 -> 8.3035 | Improved |
| add PCA and decoder warm-up to the multi-head model (ablation) | UniTalker · multi-dataset training stability | training convergence | Improved |
| add pivot identity embedding (ablation) | UniTalker · cross-ID inference | face motion quality | Improved |
| remove autoregression from FaceFormer (ablation) | FaceFormer · BIWI-Test-A | inference time (s): 0.705 -> 0.024 | Improved |
| use TCN instead of transformer as motion decoder (ablation) | UniTalker-L · A2F-Bench D0-D7 | LVE: 3.971 -> 3.859 (D0) | Mixed |
| place frequency adaptor behind the transformer encoder (ablation) | UniTalker-L · A2F-Bench D0-D7 | LVE: 3.951 -> 3.859 (D0) | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- FaceFormer — FaceFormer: Speech-Driven 3D Facial Animation With Transformers (CVPR 2022) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Probabilistic Speech-Driven 3D Facial Motion Synthesis: New Benchmarks Methods and Applications (CVPR 2024 · similarity 0.89)
- Wav2Sem: Plug-and-Play Audio Semantic Decoupling for 3D Speech-Driven Facial Animation (CVPR 2025 · similarity 0.89)
- Let's Chorus: Partner-aware Hybrid Song-Driven 3D Head Animation (CVPR 2025 · similarity 0.88)
- ProsodyTalker: 3D Visual Speech Animation via Prosody Decomposition (AAAI 2025 · similarity 0.88)
- Mimic: Speaking Style Disentanglement for Speech-Driven 3D Facial Animation (AAAI 2024 · similarity 0.88)
- DEEPTalk: Dynamic Emotion Embedding for Probabilistic Speech-Driven 3D Face Animation (AAAI 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
