Generating human-centric videos that preserve both visual identity and person-specific expressive behavior remains a fundamental challenge. In addition to reproducing appearance, a model must replicate the facial behaviors that characterize how a subject expresses emotion over time. However, most state-of-the-art methods condition generation on a single reference image, which contains no information about these temporal dynamics. As a result, they tend to preserve the subject's visual identity but often produce expressions with limited variation and weak subject specificity. To mitigate this issue, we introduce BEACON, a lightweight framework for person-specific video generation that produces more expressive videos by disentangling visual identity from expressive behavior. BEACON conditions generation on two complementary signals: a reference image encoding the identity and a reference video capturing subject-specific facial dynamics. By conditioning on these complementary signals, BEACON generates videos that better preserve both the subject's appearance and characteristic facial dynamics, while also supporting identity-expression transfer. Our experiments on the MEAD and RAVDESS datasets show that by fine-tuning on approximately 2,000 pairs and updating about 1% of the pretrained Wan video diffusion model, BEACON improves facial expressivity over state-of-the-art video generation methods while maintaining competitive identity preservation.
Figures & tables
Figure 1: Given a reference image encoding visual identity, a reference video capturing subject-specific facial dynamics, and a text prompt, BEACON generates expressive videos that preserve both the subject’s appearance and characteristic facial behaviour (top), while enabling identity-expression transfer across different subjects (bottom).
Figure 2: Overview of the proposed BEACON architecture. It conditions T2V generation on an identity reference image and an expressive reference video in addition to the text prompt. The expressive branch extracts facial dynamics from the reference video and injects them into the frozen video generation backbonwhile preserving their decoupling from visual identity.
Figure 3: BEACON training scheme. Training combines real MEAD pairs, where the identity image, expressive video, and target video all belong to the same subject, with synthetic cross-identity pairs, where the expressive video contains a different identity and the target video is generated using a Portrait Animation Model [ 21 ] .
Method
E-FID ↓
AUE ↓
FaceSim ↑
Subject Consistency ↑
Aesthetic Quality ↑
Motion Smoothness ↑
Per Subject
Per Emotion
Stand-In [ 22 ]
2.28
2.87
0.259
0.826
0.985
0.526
0.996
ConsisID [ 24 ]
3.17
3.94
0.295
0.804
0.948
0.502
0.982
BEACON
0.86
0.89
0.142
0.760
0.991
0.500
0.997
Table 1: Quantitative comparison of S2V generation methods on the MEAD dataset.
Method
Ground Truth Frame
FaceSim ↑
Stand-In [ 22 ]
Neutral
0.826
BEACON
Neutral
0.760
Stand-In [ 22 ]
Expressive
0.678
BEACON
Expressive
0.852
Table 2: Face similarity evaluated with neutral and expressive reference frames.
Method
E-FID ↓
AUE ↓
FaceSim ↑
Subject Consistency ↑
Aesthetic Quality ↑
Motion Smoothness ↑
Per Subject
Per Emotion
Stand-In [ 22 ]
2.89
1.83
0.202
0.805
0.990
0.527
0.996
ConsisID [ 24 ]
3.97
2.85
0.248
0.802
0.958
0.523
0.991
BEACON
1.57
0.97
0.174
0.854
0.993
0.522
0.996
Table 3: Comparison of S2V generation methods on the RAVDESS dataset.
Figure 4: Qualitative comparison of generated videos. Examples are shown from two datasets: MEAD (left) and RAVDESS (right).
Figure 5: E-FID matrices between generated videos and ground-truth videos across all emotion categories on MEAD dataset. Left: Stand-In. Right: BEACON .
Method
Face Similarity
Identity Ref. ↑
Expression Ref. ↓
E-FID ↓
BEACON w/o Synthetic Data
0.10
0.67
0.83
BEACON
0.69
0.25
1.01
Table 4: Ablation of synthetic training data for identity-expression disentanglement
Figure 6: Identity-expression preservation and transfer. Examples of the different inference settings enabled by BEACON . Identity and expressive references can originate from the same subject for joint preservation, or from different subjects for identity-expression transfer.