Production-ready human video generation requires digital actors to maintain strictly consistent full-body identities across dynamic shots, viewpoints and motions, a setting that remains challenging for existing methods. Prior methods often suffer from face-centric behavior that neglects body-level consistency, or produce copy-paste artifacts where subjects appear rigid due to pose locking. We present Actor-18M, a large-scale human video dataset designed to capture identity consistency under unconstrained viewpoints and environments. Actor-18M comprises 1.6M videos with 18M corresponding human images, covering both arbitrary views and canonical three-view representations. Leveraging Actor-18M, we propose WildActor, a framework for any-view conditioned human video generation. We introduce an Asymmetric Identity-Preserving Attention mechanism coupled with a Viewpoint-Adaptive Monte Carlo Sampling strategy that iteratively re-weights reference conditions by marginal utility for balanced manifold coverage. Evaluated on the proposed Actor-Bench, WildActor consistently preserves body identity under diverse shot compositions, large viewpoint transitions, and substantial motions, surpassing existing methods in these challenging settings.
Figures & tables
Figure 2: Construction pipeline and representative samples of Actor-18M . Left: Frames are sampled and filtered from identity-consistent videos, from which facial and body images are extracted as ground-truth references. Right: Based on these references, view-transformed samples are generated to construct Actor-18M -A , while attribute-conditioned image editing under diverse environments, lighting conditions, and motions produces Actor-18M -B . Canonical three-view images in Actor-18M -C are generated using Nano-Banana, guided by frames with the highest visibility selected from different viewpoints, serving as complete identity anchors.
Subset
Region
Source
Quantity
Viewpoint Distribution (%)
A
Face
Self-Crop
1.05M
F:30.8 / L:16.3 / R:25.0 / U:2.9 / D:25.0
View-Aug
5.72M
F:8.6 / L:30.2 / R:40.9 / U:10.0 / D:10.3
Body
Self-Crop
1.64M
F:62.8 / S:36.6 / B:0.6
View-Aug
8.73M
F:26.2 / S:71.6 / B:2.2
B
Face
Self-Crop
149K
F:33.9 / L:18.5 / R:25.1 / U:8.2 / D:14.3
Attr-Aug
281K
F:68.4 / L:2.8 / R:21.2 / U:3.9 / D:3.7
Table 1: Detailed statistics of Actor-18M . Self-Crop : references cropped directly from raw videos. View-Aug and Attr-Aug : generated data for viewpoint and attribute augmentation, respectively. Viewpoints include Front (F), Left (L), Right (R), Up (U), Down (D) for faces, and Front (F), Side (S), Back (B) for bodies. The generated data substantially mitigates the strong frontal-view bias.
Figure 3: Overview of WildActor . Left: The overall architecture, where a video DiT is conditioned on multi-view facial and body reference images selected via Viewpoint-Adaptive Monte Carlo Sampling. Reference tokens are embedded with I-RoPE to distinguish them from video tokens in the shared spatio-temporal attention space. Right: Details of AIPA, illustrating how identity reference tokens are aggregated and injected into video tokens through asymmetric attention while preserving prior of backbone.
Figure 4: Qualitative comparison on sequential narrative. WildActor maintains stronger full-body consistency and prompt adherence than prior methods under viewpoint changes, camera motion, and scene transitions. Zoom in to better compare fine-grained details.
Method
Params
Face Identity ↑
Body Consistency ↑
Semantic Alignment
Feature-Level ↑
VLM-Level ↑
Setting 1: Sequential Narrative
Qwen-Image-Edit + I2V
5B
0.521
0.720
0.228
0.733
T2V → I2V ( w/o Ref)
5B
0.320
0.450
0.215
0.613
T2V → I2V ( w Ref, Ours)
5B
0.548
0.925
0.235
0.893
Setting 2: Contextual Generalization
Table 2: Quantitative comparisons on Actor-Bench under (1) Sequential Narrative and (2) Contextual Generalization . Best and runner-up are bold and underlined . ∗ denotes closed-source commercial models.
Figure 6: Ablation study on data strategies and model components. We evaluate variants in controlled scenes with turning motions, enabling clear comparison under viewpoint changes.
Setting
Body Consistency ↑
Front
Side
Back
Average
Raw-Crop
0.885
0.725
0.680
0.802
Random Sampling
0.915
0.840
0.785
0.865
Viewpoint-Adaptive
0.958
0.952
0.937
0.952
Table 3: Ablation of dataset strategies. We report Body Consistency scores across different target viewpoints.
Setting
Component
Face ID ↑
Body Cons. ↑
Semantic Alignment
AIPA
I-RoPE
Feature-Level ↑
VLM-Level ↑
Full-Attn
✗
✓
0.515
0.890
0.215
0.610
w/ AIPA only
✓
✗
0.542
0.825
0.230
0.895
WildActor
✓
✓
0.559
0.952
0.238
0.920
Table 4: Ablation of model components. We analyze the impact of AIPA and I-RoPE on identity and semantic alignment.