Cross-species portrait animation, which aims to animate a source portrait of species A using a driving video of species B, is particularly challenging due to substantial anatomical differences in facial structures across species. Previous works typically fine-tune human portrait animation models on animal video datasets but often suffer from driving-species leakage. To address this, we propose CrossPortrait, a novel framework for expressive cross-species portrait animation, supporting diverse scenarios including human-to-animal, animal-to-human and cross-species animal-to-animal (e.g. cat-to-dog) animation. Specifically, we decompose the portrait motion into rigid head poses and expression keypoints. To accommodate species-specific facial structures while facilitating cross-species expression alignment, we introduce a species-routed expression encoder that incorporates species-specific LoRAs into a unified expression encoder. Furthermore, we introduce expression-level adversarial domain adaptation (Ex-ADA) to suppress driving-species-specific information in the expression latents, thereby mitigating species leakage and promoting a shared expression space across species. Extensive experiments show that our method enables expressive and identity-preserving portrait animation across diverse species and outperforms other state-of-the-art approaches.
Figures & tables
Figure 1: Cross-species animation by CrossPortrait. Our method transfers motion while preserving source identity in human-to-animal, animal-to-animal, and animal-to-human scenarios. Blue , red , and green borders denote source portraits, driving images, and CrossPortrait outputs , respectively.
Figure 2: Comparison of species leakage with LivePortrait ( Guo et al., 2024 ) , FaceShot ( Gao et al., 2025 ) , and Wan-Animate-2 ( Wang et al., 2026 ) . These approaches tend to transfer human dental characteristics to the animated cat, whereas our method faithfully follows the driving mouth motion while preserving the source cat’s oral morphology. For a fair comparison, we further post-train all applicable baselines on the same animal training data.
Figure 3: Stage 1: Learning cross-species motion alignment. We align expression representations across species while suppressing species-specific identity cues through the species-routed expression encoder and expression-level adversarial domain adaptation (Ex-ADA).
Figure 4: Stage 2: DiT-based cross-species portrait animation. We inject the pose and expression features extracted from pretrained pose and expression encoders into the video DiT for expressive cross-species animation.
Method
Animal Self-Reenactment
H → A User Preference: Method over CrossPortrait
PSNR ↑
SSIM ↑
LPIPS ↓
L1↓
Expr. (%) ↓
Pose (%) ↓
ID (%) ↓
Overall (%) ↓
Follow-Your-Emoji
21.38
0.7639
0.1730
0.05030
0.0
0.0
0.0
0.0
FaceShot
16.17
0.6019
0.3822
0.1012
3.2
1.6
12.9
3.2
LivePortrait-Animal
24.37
0.7720
0.1438
0.04010
21.2
21.2
34.6
23.1
FantasyPortrait
N/A
N/A
N/A
N/A
0.0
0.0
12.9
0.0
Wan-Animate-2
23.54
0.7834
0.1955
0.0590
30.0
22.5
34.2
19.8
Table 1: Animal self-reenactment reconstruction and H → A pairwise preference evaluation. Self-reenactment uses 50 cat and 50 dog videos. Each preference value denotes the rate at which users prefer the corresponding method over CrossPortrait .
Method
A → H (Balanced cat/dog drivers)
Cross-species A → A (A1 → A2)
CSIM ↑
Expr. (%) ↓
Pose (%) ↓
ID (%) ↓
Overall (%) ↓
Expr. (%) ↓
Pose (%) ↓
ID (%) ↓
Overall (%) ↓
FaceShot
0.8657
0.0
0.0
6.2
0.0
0.0
0.0
0.0
0.0
Follow-Your-Emoji
0.8445
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
LivePortrait-Animal
0.8625
0.0
0.0
0.0
0.0
0.0
0.0
12.5
0.0
Wan-Animate-2
0.8629
0.4
8.8
4.2
7.9
6.2
8.2
0.0
3.8
No-Alignment Baseline
0.8451
18.2
8.2
0.0
12.2
23.9
17.1
10.9
20.6
Table 2: A → H and cross-species A → A (A1 → A2) evaluation. A → H uses balanced cat and dog driving clips. Each value denotes the preference rate for the corresponding method over CrossPortrait ; lower values indicate better performance by CrossPortrait .
Figure 5: Qualitative comparison of CrossPortrait with competing methods. The top, middle, and bottom rows show human-to-animal, animal-to-human, and cross-species cat ↔ dog animation, respectively. Blue , red , and green borders indicate source portraits, driving images, and CrossPortrait outputs , respectively. Red boxes mark visible artifacts, and gray cells indicate unsupported cases.
Figure 6: Qualitative H → A ablations. Left: No-Alignment and w/o expression-level ADA. Right: w/o species-routed expression encoder.
Variant
Expr. (%) ↓
Pose (%) ↓
ID (%) ↓
Overall (%) ↓
(A) No-alignment baseline
6.4
15.4
11.5
14.1
(B) w/o Ex-ADA
23.1
26.9
23.1
19.2
(C) w/o species-routed expression encoder
21.2
30.3
33.3
27.9
(D) w/o temporal stabilization.
30.0
48.0
44.0
32.0
Table 3: Ablation user study on bidirectional human–animal animation. Each value denotes the preference rate for the corresponding method over CrossPortrait ; lower values indicate better performance by CrossPortrait .
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: Pose branch for dense rigid pose-flow. PoseNet predicts pitch, yaw, roll, translation, and scale; these parameters are converted into the dense three-channel displacement field Δpose .
Method
Self-Reenactment
Cross-Identity Reenactment
PSNR ↑
SSIM ↑
LPIPS ↓
L1↓
CSIM ↑
AED ↓
APD ↓
CSIM ↑
Follow-Your-Emoji
22.20
0.7650
0.1204
0.0428
0.8130
0.3039
0.0540
0.4282
LivePortrait-Human
20.83
0.7419
0.1443
0.0518
0.8750
0.3328
0.2615
0.5528
FaceShot
17.99
0.6836
0.2382
0.0696
0.7078
0.3868
0.3104
0.2971
HunyuanPortrait
21.46
0.7584
0.1328
0.0459
0.8589
0.2270
0.1182
0.7208
X-NeMo
21.12
0.7501
0.1397
0.0486
0.8463
0.2250
0.0200
0.6650
Appendix
Table 4: Human-to-human sanity check. Self-reenactment is evaluated with pixel-level ground truth, while cross-identity reenactment uses unpaired human motion and identity metrics.
Figure 8: Qualitative H → H cross-identity comparison on fixed source–driving pairs. Columns show the source portrait, driving frame, and outputs from CrossPortrait , Wan-Animate, X-NeMo, Follow-Your-Emoji, FaceShot, LivePortrait-Human, and HunyuanPortrait. Red boxes highlight local artifacts in the eye or mouth regions.
Method
H → H
Same-species A → A
H → A
A → H
Cross-species A → A
Cross-domain and animal-capable baselines
Follow-Your-Emoji
Yes
Partial
Yes
Partial
Partial
LivePortrait-Animal
N/A
Yes
Yes
Yes
Partial
FaceShot
Yes
Yes
Yes
Yes
Yes
FantasyPortrait
Yes
N/A
Yes
N/A
N/A
Wan-Animate-2
Yes
Yes
Yes
Yes
Yes
Appendix
Table 5: Applicability of compared methods under our evaluation protocol.
Method
Valid n
Expr. (%) ↑
Pose (%) ↑
ID (%) ↑
Overall (%) ↑
FaceShot
50
100
100
100
100
Follow-Your-Emoji
50
100
100
100
100
LivePortrait-Animal
50
100
100
91
100
Wan-Animate-2
50
100
98
100
100
Appendix
Table 6: Same-species animal cross-identity user study. Entries are pairwise preference rates for CrossPortrait over each evaluated baseline, with ties counted as half a vote. The preference columns are percentages (%): Expr. measures expression fidelity and temporal alignment; Pose measures head-pose agreement; ID measures source identity and morphology preservation, including driving-species leakage; and Overall measures motion, source preservation, temporal coherence, and visual quality.
Figure 9: Qualitative unseen-source generalization by CrossPortrait . Our method transfers human driving motion while preserving the identity and species-specific morphology of source portraits from unseen species. Blue , red , and green borders denote source portraits, driving images, and CrossPortrait outputs , respectively. The driving species is human and is covered by training; this figure therefore evaluates source-side generalization only and does not evaluate unseen-species driving.
School of Computer Science and Engineering, Southeast University, Nanjing, China · Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications, Ministry of Education, China