Animating a reference image with a driver's motion requires a motion representation that captures face, hand, and body articulation but excludes the driver's appearance. Most body animation methods take motion from external pose estimators and render it with many denoising steps; compact learned motion representations mostly cover the face. We present FUHA, a real-time-capable feed-forward transformer that learns a compact motion embedding of the whole human and uses it for face, upper-body, and full-body animation. A Q-Former compresses the driver's features from a frozen visual encoder into this embedding. To disentangle motion from appearance, a novel contrastive loss pulls it toward the motion embedding of the driver's rendered keypoint image and away from that of the same person in another pose. At inference time, FUHA needs no external pose estimators, keypoints, or parametric body models. As a retrieval key, the motion embedding matches body pose better than the learned global embeddings of ten vision encoders. Its same-person retrieval rate is 0.1%, against 9.5% for its own DINOv3 backbone. On our upper-body benchmark, FUHA outperforms all baselines on every self-reenactment metric and on cross-reenactment identity similarity. It is competitive on face and full-body and runs at up to 38.9 FPS at 512x512 on a single RTX 4090.
Figures & tables
Figure 1: FUHA drives face, upper-body, and full-body animation from one compact motion embedding. ( 1(a) ) Method categories against the three requirements; no prior category meets all three. ( 1(b) ) Speed vs. cross-reenactment expression error per benchmark; FUHA-C/T (Cosmos/TAESD3 tokenizers) appear in both.
Figure 2: FUHA inference flow. Given a reference image and a driver frame, two pretrained, frozen backbones (an RGB tokenizer and a visual feature extractor) produce the cross-attention inputs for two trainable Q-Formers. The Appearance Encoder sums RGB tokens and features and compresses them via cross-attention into a fixed set of appearance embeddings; the Motion Encoder attends to visual features only, yielding a compact motion embedding. A Motion Decoder expands it into a motion map, and the Appearance Decoder cross-attends motion to appearance to produce the final output embeddings, which the frozen RGB tokenizer decodes back to an image. The entire pipeline is feed-forward, frame-by-frame, and real-time-capable.
Figure 3: Contrastive loss. The driver’s motion embedding is pulled toward the rendered keypoints’ and away from the reference’s.
Method
Self-Reenactment
Cross-Reenactment
L1 ↓
SSIM ↑
LPIPS ↓
ID ↑
AED ↓
AKD-B ↓
AKD-H ↓
FID ↓
FVD ↓
ID ↑
AED ↓
AKD-B ↓
AKD-H ↓
FID ↓
FVD ↓
Moore-AA ( Hu et al., 2024 )
0.0619
0.694
0.206
0.657
0.548
42.3
81.7
26.2
185.9
0.511
0.848
56.5
103.1
53.2
368.0
Champ ( Zhu et al., 2024 )
0.0980
0.611
0.335
0.571
0.670
56.3
103.2
33.8
274.8
0.452
0.910
60.2
117.8
56.4
378.8
MimicMotion ( Zhang et al., 2025b )
0.0816
0.677
0.278
0.582
0.506
43.2
65.6
27.1
80.3
0.131
0.702
53.8
86.1
64.8
203.7
Wan-Animate ( Cheng et al., 2025 )
0.0643
0.712
0.219
0.668
0.467
37.6
70.5
24.4
133.4
0.489
0.765
50.8
93.4
57.5
207.4
UniAnimate-DiT ( Wang et al., 2025b )
0.0717
0.724
0.201
0.694
0.618
46.6
86.6
20.3
66.0
0.630
0.850
55.7
97.2
50.2
145.8
Table 1: Upper-body animation evaluation on our benchmark at 5122 ( 1st , 2nd , 3rd ).
Figure 4: Qualitative results for upper-body cross-reenactment. FUHA keeps the reference identity while following the driver’s expression, body pose, and hand pose. The final two columns show FUHA-C and FUHA-T, respectively. More results, including face and full-body, are in Appendix B .
Method
Self-Reenactment
Cross-Reenactment
L1 ↓
SSIM ↑
LPIPS ↓
ID ↑
AED ↓
APD ↓
FID ↓
FVD ↓
ID ↑
AED ↓
APD ↓
FID ↓
FVD ↓
LivePortrait ( Guo et al., 2024 )
0.0990
0.655
0.359
0.868
0.341
0.0256
40.4
224.7
0.658
0.885
0.0705
164.9
533.0
X-Portrait ( Xie et al., 2024 )
0.0820
0.685
0.271
0.855
0.414
0.0449
37.1
271.3
0.793
0.974
0.1186
162.9
655.5
FollowYE ( Ma et al., 2024 )
0.0746
0.699
0.250
0.844
0.440
0.0282
39.4
184.6
0.747
1.012
0.0483
160.0
678.2
X-NeMo ( Zhao et al., 2025 )
0.1072
0.600
0.376
0.832
0.401
0.0206
43.8
245.4
0.654
0.818
0.0486
154.3
546.6
HunyuanPortrait ( Xu et al., 2025 )
0.0943
0.660
0.312
0.853
0.376
0.0313
35.0
203.8
0.691
0.894
0.0880
159.8
571.8
Table 2: Face animation evaluation on the TalkingHead-1KH ( Wang et al., 2021 ) test split for self-reenactment. For cross-reenactment we use 100 FFHQ ( Karras et al., 2019 ) images as reference, paired randomly with a TalkingHead-1KH test set clip. All methods receive FFHQ-aligned 5122 frames. Table cells are shaded based on per-column rank: 1st , 2nd , and 3rd .
Method
L1 ↓
SSIM ↑
PSNR ↑
LPIPS ↓
ID ↑
AED ↓
AKD-B ↓
AKD-H ↓
FID ↓
FVD ↓
Moore-AA ( Hu et al., 2024 )
0.0812
0.770
17.58
0.302
0.448
0.733
48.6
87.5
44.8
235.8
Champ ( Zhu et al., 2024 )
0.1223
0.719
14.45
0.399
0.356
0.993
85.3
118.9
42.2
281.1
MimicMotion ( Zhang et al., 2025b )
0.1118
0.728
15.40
0.363
0.282
0.778
65.8
88.9
37.9
188.7
Animate-X ( Tan et al., 2025 )
0.1145
0.745
15.00
0.385
0.421
0.961
91.3
151.2
31.0
229.5
Wan-Animate ( Cheng et al., 2025 )
0.0815
0.768
17.76
0.310
0.427
0.705
46.8
84.5
36.1
168.2
UniAnimate-DiT ( Wang et al., 2025b )
0.0672
0.818
19.11
0.247
0.458
0.705
44.9
76.4
25.9
170.8
Table 3: Full-body animation evaluation on the TikTok ( Jafarian & Park, 2021 ) test set (self-reenactment). ( 1st , 2nd , 3rd ).
Self-Reenactment
Cross-Reenactment
Variant
LPIPS ↓
ID ↑
FID ↓
FVD ↓
ID ↑
AED ↓
AKD-B ↓
AKD-H ↓
FID ↓
FVD ↓
1 motion token
0.351
0.134
78.8
258.5
0.088
0.724
57.8
92.6
88.0
310.8
2 motion tokens
0.106
0.820
15.2
38.9
0.701
0.686
51.5
88.8
55.9
171.7
4 motion tokens
0.127
0.812
17.7
63.1
0.733
0.761
58.9
91.6
55.7
201.9
w/o contrastive loss
0.340
0.199
88.8
263.9
0.057
0.606
47.8
80.3
101.1
325.9
w/o DINO in appearance
0.125
0.811
18.3
59.4
0.712
0.732
55.8
91.1
56.1
202.4
Table 4: FUHA-C ablations on the upper-body set at 5122 . Table C.1 adds a ReLU variant of the contrastive loss. ( 1st , 2nd , 3rd ).
Figure 5: Motion latent-space walk. We linearly interpolate between the FUHA-C motion embeddings of two drivers and decode each step on one reference. Over 25 measured pose and expression walks, motion tracks t (median Spearman ρ=1.00 , R2=0.99 ) and identity stays with the reference (median ArcFace 0.75; at most 0.19 to either driver; Appendix A.2 ).
Table 10
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
Encoder
AKD-B ↓
AKD-H ↓
Head pose ↓
Same person ↓
(px)
(px)
(degrees)
(%)
Random
110.8
166.2
10.61
0.0
SigLIP2 base-p16-224
91.0
157.1
10.23
10.4
SigLIP2 so400m-p16-512
91.3
152.4
10.09
11.3
Sapiens 0.3B
98.9
167.9
10.22
3.2
Sapiens 1B
102.3
162.1
10.35
3.8
Appendix
Table A.1: Motion retrieval with learned global embeddings. Body/hand keypoint errors (AKD-B/H) are in pixels at 5122 ; head-pose error is in degrees. Pose scores use a person-disjoint gallery; same-person retrieval is measured separately with shared identities retained. Lower is better. AKD-H uses 475 queries with visible hands.
Figure A.1: Motion retrieval across identities. Selected queries (left) and their nearest neighbors from FUHA-C and five vision-encoder baselines. The gallery excludes the query’s identity. Baseline variants are DINOv3 ViT-B/16, C-RADIOv4-SO400M, Sapiens2-0.1B, Sapiens-0.3B, and SigLIP2-base-p16-224, from left to right.
Figure A.2: Walk measurements over 25 walks and six references. Left: dominant SMIRK parameter, normalized by the decoded endpoints (thin: walk and reference; thick: median; dotted: linear). Right: ArcFace similarity to the reference (solid: median; band: 10th–90th percentile) and to either driver (dashed: maximum). Same: both drivers from one clip; cross: from two people.
Figure A.3: Motion latent-space walks. We linearly interpolate between the FUHA-C motion embeddings of two driver frames (left and right) and decode each step on two fixed references. The top three walks change head pose and expression; the bottom two move the hands. Walks 3 and 5 interpolate between two different people. Motion changes smoothly while the reference identity is preserved.
(a) Upper-body
Setting
Model
L1 ↓
SSIM ↑
LPIPS ↓
ID ↑
AED ↓
AKD-B ↓
AKD-H ↓
FID ↓
FVD ↓
Self
Wan-Animate
0.0643
0.712
0.219
0.668
0.467
37.6
70.5
24.4
133.4
UniAnimate-DiT
0.0717
0.724
0.201
0.694
0.618
46.6
86.6
20.3
66.0
FUHA-C
0.0287
0.838
0.106
0.820
0.315
27.1
48.5
15.2
38.9
FUHA-W
0.0413
0.781
0.140
0.781
0.472
32.8
56.6
18.4
27.9
Cross
Wan-Animate
–
–
–
0.489
0.765
50.8
93.4
57.5
207.4
Appendix
Table A.2: Animation with a diffusion decoder. FUHA-W uses the FUHA-C motion embedding to condition Wan. Baseline results are reproduced from Tables 1 – 3 . Cells are ranked within each setting and region ( 1st , 2nd , 3rd ).
Figure A.4: Qualitative results with FUHA-W. The diffusion decoder used in FUHA-W yields improved visual details at a slight expense of pose accuracy. All methods use the same reference image and driving frame in each row.
Method
Face
Upper-body
Full-body
LivePortrait
1.96
–
–
X-Portrait
2.60
–
–
FollowYE
1.90
–
–
X-NeMo
1.97
–
–
HunyuanPortrait
2.01
–
–
PersonaLive
2.47
–
–
Appendix
Table B.1: Temporal consistency : self-reenactment tLP ↓ ( ×100 ). – marks methods not evaluated on a benchmark. Cells are shaded by per-column rank: 1st , 2nd , and 3rd .
Figure B.1: Qualitative results for face animation. The final two columns show FUHA-C and FUHA-T, respectively. Beyond head pose and facial expression, FUHA captures shoulder pose and camera-relative head distance.
Figure B.2: Additional qualitative results for upper-body animation. The final two columns show FUHA-C and FUHA-T, respectively.
Figure B.3: Qualitative results for full-body animation. The final two columns show FUHA-C and FUHA-T, respectively. FUHA transfers body pose and facial expression accurately.
Self-Reenactment
Cross-Reenactment
Variant
LPIPS ↓
ID ↑
FID ↓
FVD ↓
ID ↑
AED ↓
AKD-B ↓
AKD-H ↓
FID ↓
FVD ↓
1 motion token
0.351
0.134
78.8
258.5
0.088
0.724
57.8
92.6
88.0
310.8
2 motion tokens
0.106
0.820
15.2
38.9
0.701
0.686
51.5
88.8
55.9
171.7
4 motion tokens
0.127
0.812
17.7
63.1
0.733
0.761
58.9
91.6
55.7
201.9
ReLU contrastive
0.125
0.813
16.8
54.6
0.730
0.709
55.4
91.9
55.6
189.2
w/o DINO in appearance
0.125
0.811
18.3
59.4
0.712
0.732
55.8
91.1
56.1
202.4
Appendix
Table C.1: FUHA-C ablations on the upper-body set at 5122 . ( 1st , 2nd , 3rd ).
Figure C.1: Qualitative ablation comparisons on the upper-body cross-reenactment set. Each row is a (reference, driver) pair; from left to right we show the reference image, the driver frame, four component ablations, and the full FUHA-C model. Without the contrastive loss (third column), outputs take the driver’s hair, gender, and face shape at low detail, but the reference’s background and clothing; the other ablations stay close to FUHA-C. See Table C.1 for matching numerical values.
Human image animation aims to generate a video from a static reference image, guided by pose information extracted from a driving video. Existing approaches often rely on pose estimators to extract intermediate representations, but such signals are prone to errors under occlusion or complex poses. Building on these observations, we present DirectAnimator, a framework that bypasses pose extraction and directly learns from raw driving videos. We introduce a Driving Cue Triplet consisting of pose, face, and location cues that captures motion, expression, and alignment in a semantically rich yet stable form, and we fuse them through a CueFusion DiT block for reliable control during denoising. To make learning dependable when the driving and reference identities differ, we devise a Same2X training strategy that aligns cross-ID features with those learned from same-ID data, regularizing optimization and accelerating convergence. Extensive experiments demonstrate that DirectAnimator attains state-of-the-art visual quality and identity preservation while remaining robust to occlusions and complex articulation, and it does so with fewer computational resources. Our project page is at https://directanimator.github.io/.
Yuan Zeng, Yujia Shi, Yuhao Yang +4
1Tsinghua University · 2Harbin Institute of Technology · 3Pengcheng Laboratory
Avatar reconstruction has traditionally relied on per-subject optimization that requires hours of computation or on expensive preprocessing that limits scalability. We introduce FFAvatar, a generalizable feed-forward framework that reconstructs high-quality, animatable 3D Gaussian head avatars from few-shot unposed portrait images in seconds. FFAvatar fuses information from multiple source images into a unified canonical Gaussian representation through Multi-View Query-Former, which is animated via FLAME parameters predicted end-to-end directly from pixels, eliminating the overhead of offline FLAME extraction. We further propose a three-stage training curriculum that achieves both broad generalization and high-fidelity reconstruction: (i) scalable pretraining on extensive monocular video data with over 1M identities to learn strong generalizable priors; (ii) multi-view fine-tuning on a small but high-quality dataset of 360-degree captures to enhance geometric fidelity and extreme-view awareness; and (iii) optional personalization that adapts to specific identities for maximum fidelity within 500 optimization steps. Extensive experiments demonstrate that FFAvatar sets a new standard for identity preservation, geometric consistency, and animation fidelity. On the NeRSemble benchmark, it outperforms the state-of-the-art LAM by a substantial 5.5 PSNR gain. Furthermore, FFAvatar enables real-time deployment, reconstructing avatars in 2 seconds without personalization and 10 seconds with personalization, while supporting 49 FPS animation on a single NVIDIA A100 GPU.
Thuan Hoang Nguyen, Jiahao Luo, Yinyu Nie +3
Snap Inc. · MBZUAI · University of California, Santa Cruz
We present FFAvatar, a Transformer-based 3D Gaussian framework for fast construction of high-quality and animatable 4D head avatars from one or more reference portrait images. Unlike existing feed-forward approaches that require a fixed number of input views, FFAvatar supports incremental reconstruction, progressively refining the avatar representation as additional reference images become available. At the core of our method is an alternating attention mechanism that disentangles identity appearance from expression and viewpoint variations, enabling the reconstruction of a canonical 3D appearance that remains consistent across poses and facial expressions. To balance visual fidelity and computational efficiency, we introduce a sparse-to-dense learning paradigm. Coarse appearance features are first learned using sparse primitives anchored to the FLAME vertex level and are subsequently densified in the UV domain to capture fine-grained geometric and texture details. We further propose a plug-and-play motion refinement module that enables subject-specific dynamic personalization by modeling residual motion beyond parametric deformation. Extensive experiments demonstrate that FFAvatar efficiently produces high-fidelity and controllable 4D head avatars, achieving superior flexibility, driving efficiency, and identity-consistent rendering across diverse expressions and viewpoints.
Jianjiang Yao, Ke Xian, Renxiang Dai +1
School of Electronic Information and Communications, Huazhong University of Science and Technology, Wuhan 430074, China