Organizations: Media Integration and Communication Center (MICC), University of Florence, Italy · University of Copenhagen · IMT Nord Europe, Institut Mines-Télécom, Centre for Digital Systems · Univ. Lille, CNRS, Centrale Lille, Institut Mines-Telecom, UMR 9189 CRIStAL, F-59000 Lille, France
Speech-driven 3D facial animation has advanced rapidly, yet most approaches remain tied to registered template meshes, preventing effective deployment on raw 3D scans with arbitrary topology. At the same time, modeling controllable emotional dynamics beyond lip articulation remains challenging, and is often tied to template-based parameterizations. We address these challenges by proposing FreeTalk, a two-stage framework for emotion-conditioned 3D talking-head animation that generalizes to unregistered face meshes with arbitrary vertex count and connectivity. First, Audio-To-Sparse (ATS) predicts a temporally coherent sequence of 3D landmark displacements from speech audio, conditioned on an emotion category and intensity. This sparse representation captures both articulatory and affective motion while remaining independent of mesh topology. Second, Sparse-To-Mesh (STM) transfers the predicted landmark motion to a target mesh by combining intrinsic surface features with landmark-to-vertex conditioning, producing dense per-vertex deformations without template fitting or correspondence supervision at test time. Extensive experiments show that FreeTalk matches specialized baselines when trained in-domain, while providing substantially improved robustness to unseen identities and mesh topologies. Code and pre-trained models will be made publicly available.
Figures & tables
Figure 1: FreeTalk takes an unregistered static face mesh (in blue on the left) and an emotional audio speech and predicts a sequence of deformation fields (in cyan ) matching speech-related motions and expression-related deformations.
Emotion
Any mesh
CodeTalker / FaceFormer / SelfTalk
✗
✗
EmoTalk / Emote / Emovoca
✓
✗
ScanTalk
✗
✓
FreeTalk
✓
✓
Table 1: Comparison of talking head models.
Figure 2: Overview of the proposed approach. ATS takes the speech audio signal A and generates a sequence of landmarks displacements (ΔLt)t=1T . Then, STM takes a static neutral mesh M and maps the sparse motion sequence of landmarks (ΔLt)t=1T to dense vertex motions to predict the final animated mesh sequence (M^t)t=1T .
Figure 3: Overview of the Sparse-To-Mesh (STM) module. STM predicts vertex-wise features on a static mesh M , and landmark features from displacement vectors L . The mesh features are enhanced by a learnable mapping that injects a global embedding of the landmark displacements into each vertex via cross-attention.
LVE ↓
MVE ↓
FDD ↓
DTW ↓
DFD ↓
δM↓
δCd↓
Faceformer (2)
12.7
1.87
14.4
3.41
13.3
3.11
0.98
EmoFace (2)
12.8
1.85
14.9
3.46
13.3
3.10
0.99
FreeTalk (1,2,4,5 + 1,2,4,5)
16.4
2.29
3.35
2.65
11.4
4.61
0.83
FreeTalk (1,2,4,5 + 2)
16.6
2.31
3.52
2.65
11.4
4.60
0.83
FreeTalk (2 + 1,2,4,5)
13.7
2.27
4.93
2.61
11.1
3.23
0.82
FreeTalk (2 + 2)
13.8
2.27
5.12
2.60
11.0
3.25
0.81
Table 2: Quantitative comparison on the MEAD-EMOTE test set. We compare FreeTalk with state-of-the-art speech-driven 3D facial animation methods trained on MEAD-EMOTE. Lower values indicate better performance. Numbered settings follow the dataset order defined in Sec. 4.1 : (1) EmoVOCA, (2) MEAD-EMOTE, (3) VOCAset, (4) BIWI 6 , (5) Multiface. For FreeTalk, “ (a,b + c,d) ” denotes that ATS is trained on the datasets before “+” and STM on those after it.
Figure 4: Qualitative comparison on the MEAD-EMOTE test set. Each panel shows a sequence of rendered frames for a single identity. Left : Disgust ; right : Happy . Numbered settings follow the definition in Table 2 .
Figure 5: Results on unseen topologies.
VOCAset
BIWI 6
Multiface
LVE ↓
MVE ↓
FDD ↓
DTW ↓
DFD ↓
δM↓
δCd↓
LVE ↓
MVE ↓
FDD ↓
DTW ↓
DFD ↓
δM↓
δCd↓
LVE ↓
MVE ↓
FDD ↓
DTW ↓
DFD ↓
δM↓
δCd↓
VOCA (sd)
6.99
0.98
2.66
1.77
7.41
1.39
1.02
5.74
2.59
41.5
1.60
7.89
1.47
0.67
4.92
2.76
55.78
1.56
6.51
0.86
1.23
FaceFormer (sd)
6.12
0.93
2.16
1.33
5.39
0.86
0.58
4.08
2.16
37.1
1.56
6.61
0.87
0.55
2.45
1.45
20.2
0.87
4.13
0.22
0.73
FaceDiffuser (sd)
4.35
0.90
2.43
1.73
6.83
0.81
0.60
4.02
2.12
39.6
1.55
6.50
0.85
0.56
3.55
2.39
29.1
1.45
5.24
0.28
0.77
CodeTalker (sd)
3.55
0.89
2.26
1.33
5.66
0.80
0.55
5.19
2.64
20.6
1.50
6.61
0.91
0.59
4.09
2.38
47.9
1.44
5.95
0.55
0.96
SelfTalk (sd)
5.61
0.91
2.32
1.25
5.43
0.72
0.56
3.63
2.06
35.5
1.60
6.93
1.03
0.66
2.28
1.90
37.4
0.96
4.18
0.25
0.75
Table 3: Cross-dataset generalization results on neutral speech datasets. Lower values indicate better performance. (sd) denotes single-dataset training. Numbered settings follow the definition in Table 2 .
Feature extractor
Deformation Decoder
Shared MLP
DiffusionNet
NJF [ 1 ]
Shared MLP
0.0020
0.0023
0.0018
DiffusionNet [ 46 ]
0.0017
0.0021
0.0018
PoissonNet [ 31 ]
0.0016
0.0020
0.0017
Table 4: Component substitution analysis for the STM module. We report the Lip Vertex Error (LVE) of the STM module when trained and tested on VOCAset, BIWI and MultiFace. Each cell correspond to a different model for the mesh feature extractor (left column) and the deformation decoder (first row).
Figure 6: Vertex-wise L2 error on a predicted deformed mesh. We use different models for the deformation decoder. We highlight how a MLP as decoder tends to predict noisy deformations around the lips. On the opposite, Neural Jacobian Fields tends to smooth out "too much" to capture highly localized deformations around the right of the lower lip area.
Feature combination strategy
LVE
Concat.
0.0035
CA
0.0025
CA + Concat.
0.0023
GCN + CA + Concat. (ours)
0.0021
Table 5: Ablation studies for the feature combination component of the STM module.
Figure 7: Heatmap of the attention scores from one landmark to each mesh vertex at one frame. For a selected frame and landmark (in red on the landmark graph on the left of the figure), how strongly each mesh vertex attends to that landmark inside the model’s cross-attention module. Red values mean that the vertex relies more on the selected landmark when building its time-dependent latent feature, while blue values mean that the landmark has less influence there.