Digital characters are central to modern media, yet generating character videos with long-duration, consistent multi-view appearance and expressive identity remains challenging. Existing approaches either provide insufficient context to preserve identity or leverage non-character-centric information as the memory, leading to suboptimal consistency. Recognizing that character video generation inherently resembles an outside-looking-in scenario. In this work, we propose representing the character visual attributes through a compact set of anchor frames. This design provides stable references for consistency, while reference-based video generation inherently faces challenges of copy-pasting and multi-reference conflicts. To address these, we introduce two mechanisms: Superset Content Anchoring, providing intra- and extra-training clip cues to prevent duplication, and RoPE as Weak Condition, encoding positional offsets to distinguish multiple anchors. Furthermore, we construct a scalable pipeline to extract these anchors from massive videos. Experiments show our method generates high-quality character videos exceeding 10 minutes, and achieves expressive identity and appearance consistency across views, surpassing existing methods.
Figures & tables
Figure 2: The pipeline to construct training clips and character-centric anchor frames, e.g . , global, viewpoint, and expression. The blue arrow marks the subject’s forward orientation, whereas the green arrow marks the camera-facing direction.
Figure 3: Overall of the Gloria pipeline, which includes the source of content anchors (Superset Anchors), the manner to inject these anchors (RoPE as Weak Condition), and the overall framework with multi-modal conditions e.g . , text and audio.
Figure 3
Figure 5: Qualitative comparison of long-term, multi-view appearance, and expressive identity consistency. The results show the top-3 performing methods and our method. The red and yellow boxes highlight appearance details.
Figure 5
Long-term
Appearance & ID
Setting
Sub.
Back.
Arcf.
Setting
DINO-I
CLIP-I
Exp.
Ours
0.960
0.951
0.787
Ours
0.821
0.858
0.717
✗ Cg
0.840
0.844
0.233
✗ Cv,e
0.801
0.848
0.668
✗ SCA
0.922
0.899
0.489
✗ RWC
0.817
0.855
0.671
Table 4: Quantitative ablation results of content anchors ( Cg,v,e ) and the Superset Content Anchoring (SCA) and Rope as Weak Condition (RWC) mechanisms.
Figure 7: (a) Ablation of the global anchor ( Cg ). (b) Ablation of the ( Cv ). The top row indicates w/o the Cv , the middle row indicates w the Cv but w/o the Superset Content Anchoring (SCA), and the bottom row indicates with both. The small images at the top-left corner of each frame indicate the input Cv .
Figure 8: The left includes the first frame, viewpoint anchor frames and the text prompt. The right side shows generated videos, the bottom row indicates using “RoPE as Weak Condition” (RWC) while the top row dose not employ it.
Figure 9: Self-attention maps of the generated sequence and its attention toward the global anchor across different chunks (each lasting 5 s). The rightmost dashed column indicates the attention score from the generated sequence to the global anchor frame.
Figure 10: Left: the first frame and a back-view anchor frame. Right: the generated results and the corresponding attention score curve of the generated sequence toward the back-view anchor. The upper and lower rows show results generated under different texts.
Viewpoints
Front
Back
Left
Right
Ratio
29.6 %
17.3 %
27.9 %
25.2 %
Table 5: Ratios of different categories in viewpoint-anchor.
Expressions
Surprise
Angry
Disgust
Fear
Ratio
15.0 %
8.2 %
13.7 %
10.1 %
Expressions
Contempt
Neutral
Happy
Sad
Ratio
13.6 %
8.1 %
14.6 %
16.7 %
Table 6: Ratios of different categories in expression-anchor.
Figure 11: Qualitative results of the expression anchors. The first row shows the inputs, the second row shows the results with the expression anchor, and the third row shows the results without the anchors.
Figure 12: Subject Evaluation of (a) expressive identity consistency, and (b) multi-view appearance consistency.
Figure 13: Comparison of subject evaluation results with human-centric models across multiple evaluation dimensions.
Method
IQA ↑
AES ↑
Sync-C ↑
Sync-D ↓
FID/FVD ↓
Fantasy.
3.48
2.26
2.88
10.39
37.43 / 108.92
Huny.A.
3.40
2.28
5.17
8.37
28.39 / 93.49
WanS2V
3.56
2.33
4.99
8.40
36.10 / 93.73
Infinite.
3.47
2.31
5.20
8.29
29.56 / 76.39
Ours
3.59
2.33
5.62
8.06
25.50 / 90.86
Table 7: Quantitative comparison results on CelebV-HQ with the human-centric methods. Fantasy. denotes FantasyTalking, Huny.A. is HunyuanAvatar and Infinite. indicates InfiniteTalk.
Table 8: Performance in different model sizes (a) and different data scales (b).