Precise camera control is essential for cinematic character video generation, yet existing methods struggle to follow large-angle trajectories with combined rotation, translation, and radius variation from arbitrary initial azimuths. Such viewpoint changes also reveal character regions unseen in the initial frame, causing geometric distortions, appearance degradation, and identity drift. To systematically study these coupled challenges, we introduce CLACT, a large-scale Character-centric dataset and benchmark covering diverse Large-Angle Camera Trajectories, with accurate camera geometry and multi-view character observations for training and evaluation. We further propose WorldCast, a framework for world-consistent character video generation under diverse large-angle camera motions. Specifically, we back-project the initial frame into a 3D point cloud using estimated depth, align the point-cloud scale with the target camera trajectory, and render it along the trajectory to obtain explicit geometric guidance and coverage masks. To complement the incomplete character information, we introduce Geometry-Aligned Memory Attention (GAMA). Using relative camera poses estimated jointly from the initial frame and multi-view character anchors, we express anchor and target-view geometry in a common coordinate frame. GAMA uses this geometric alignment to guide the retrieval and injection of view-relevant character features during generation. Evaluations across diverse large-angle camera trajectories demonstrate accurate trajectory following while preserving character geometry, appearance, and identity.
Figures & tables
Figure 1: WorldCast enables accurate large-angle camera control while preserving character consistency. (a) Generated sequences along tilted circular and lemniscate camera trajectories. (b) Quantitative comparison of camera-control accuracy (ATE ↓ ), cross-view geometric consistency (MEt3R ↓ ), and image quality (IQ ↑ ). Each star denotes a method; colors distinguish method categories, while star size indicates image quality. Both axes are reversed, such that methods with lower ATE and MEt3R appear toward the upper right.
Figure 2: Overview of WorldCast . The framework combines trajectory-aligned geometric guidance with geometry-aligned multi-view character memory.
Figure 3: Overview of our training-data construction pipeline.
Camera Control
Character Consistency
Image Quality
Method
Anchors
ATE ↓
RTE ↓
RRE ↓
MEt3R ↓
Arcf. ↑
CLIP ↑
DINO ↑
Subj. ↑
IQ ↑
Camera-controlled generation
CamI2V [ 6 ]
✗
0.351
0.281
11.490
0.405
0.147
0.804
0.675
0.874
0.686
DualCamCtrl [ 11 ]
✗
0.297
0.082
3.607
0.408
0.185
0.796
0.661
0.904
0.699
GEN3C [ 8 ]
✗
0.130
0.038
1.255
0.306
0.433
0.819
0.696
0.915
0.716
Uni3C [ 9 ]
✗
0.098
0.033
1.069
0.324
0.374
0.823
0.691
0.911
0.711
Table 1: Quantitative comparison on our evaluation benchmark. ↓ and ↑ denote lower- and higher-is-better metrics, respectively. Anchors denotes the use of additional character views.
Figure 4: Qualitative comparison on large-angle camera-controlled video generation.
Camera Control
Character Consistency
Visual Quality
Variant
ATE ↓
RTE ↓
RRE ↓
MEt3R ↓
Arcf. ↑
CLIP ↑
DINO ↑
Subj. ↑
IQ ↑
w/o GAMA
0.100
0.031
1.149
0.333
0.370
0.813
0.668
0.899
0.699
w/o Anchor Geometry Alignment
0.071
0.027
1.374
0.239
0.547
0.826
0.708
0.909
0.716
w/o Anchor Masking
0.078
0.025
1.110
0.247
0.546
0.827
0.710
0.908
0.709
WorldCast (Ours)
0.065
0.020
0.896
0.228
0.555
0.830
0.719
0.917
0.719
Table 2: Ablation study of the proposed components. We remove the entire GAMA branch, anchor geometry alignment, or anchor foreground masking while keeping the remaining settings unchanged.
Camera Control
Variant
ATE ↓
RTE ↓
RRE ↓
w/o UE5 data
0.359
0.156
4.295
w/o Plücker conditioning
0.275
0.056
2.113
Full Stage 1
0.100
0.031
1.149
Table 3: First-stage camera-control ablation.
Figure 5: Effect of the number of character anchors on character consistency.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Figure B.1 : Training-data composition and camera configurations. (a) Sizes and training roles of the UE5 and Blender collections. (b) Character and environment counts shared by both collections, together with the number of UE5 character actions. (c) A representative Blender batch containing 36 rotational, 20 translational, and 16 combined trajectories. (d) Angular-span configurations of the camera trajectories.
Configuration
UE5
Blender
Training samples
50,000
30,000
Training stage
Large-angle camera control
Multi-view memory injection
Anchor conditioning
Not used
Front, back, left, and right views
Camera paths
Arcs, orbits, orbit-and-dolly, and free-form paths
Orbits, translations, and combined rotation–translation paths
Motion variation
Angular span, camera distance, and speed
Initial azimuth, angular span, translation, and orbit radius
Appendix
Table B.1 : Comparison of the UE5 and Blender training collections.
Figure D.1 : View-aware anchor routing under a 360∘ camera orbit. Attention responses are visualized for the 21 temporal latent positions corresponding to an 81 -frame generated video. Stars denote the geometrically expected observations with the closest camera poses to the target viewpoints, while outlined cells indicate the observations receiving the strongest attention responses. The predicted routing agrees with the expected routing for 19/21 latent positions.
Figure D.2 : Effect of Anchor Geometry Alignment on attention routing. Routing accuracy during training with and without anchor geometry alignment. Geometry alignment enables faster learning of view-aware anchor retrieval and achieves higher routing accuracy after convergence.
Figure D.3 : Additional qualitative comparison with state-of-the-art methods under diverse large-angle camera motions.
Figure D.4 : Additional qualitative comparison under large-angle camera motion with different initial viewpoints.
Figure D.5 : Additional qualitative comparison on combined rotation and translation trajectories.
Figure D.6 : Additional qualitative comparison on complex large-angle camera trajectories.
Figure D.7 : Additional qualitative comparison across diverse camera motions and initial azimuths.
Figure D.8 : Trajectory generalization using the same set of character anchors across diverse target camera motions.
Figure D.9 : Trajectory generalization using the same set of character anchors across diverse target camera motions.
For artistic applications, video generation requires fine-grained control over both performance and cinematography, i.e., the actor's motion and the camera trajectory. We present ActCam, a zero-shot method for video generation that jointly transfers character motion from a driving video into a new scene and enables per-frame control of intrinsic and extrinsic camera parameters. ActCam builds on any pretrained image-to-video diffusion model that accepts conditioning in terms of scene depth and character pose. Given a source video with a moving character and a target camera motion, ActCam generates pose and depth conditions that remain geometrically consistent across frames. We then run a single sampling process with a two-phase conditioning schedule: early denoising steps condition on both pose and sparse depth to enforce scene structure, after which depth is dropped and pose-only guidance refines high-frequency details without over-constraining the generation. We evaluate ActCam on multiple benchmarks spanning diverse character motions and challenging viewpoint changes. We find that, compared to pose-only control and other pose and camera methods, ActCam improves camera adherence and motion fidelity, and is preferred in human evaluations, especially under large viewpoint changes. Our results highlight that careful camera-consistent conditioning and staged guidance can enable strong joint camera and motion control without training. Project page: https://elkhomar.github.io/actcam/.
Omar El Khalifi, Thomas Rossi, Oscar Fossey +6
Kinetix, France · University of Oxford, United Kingdom · MBZUAI, United Arab Emirates
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.
Yuhang Yang, Fan Zhang, Huaijin Pi +5
MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, USTC · UNSW · HKU +1
Camera-controlled video generation has achieved remarkable progress in recent years. However, existing video-to-video re-rendering methods primarily rely on Supervised Fine-Tuning using synthetic datasets. At present, there is an extreme scarcity of synchronized, multi-view real-world video data. Consequently, the prevailing paradigm often exhibits limited generalization when processing out-of-distribution real-world videos, with models struggling to accurately adhere to physical scales and camera trajectories. To bridge this gap, we propose Geo-Align, the first Reinforcement Learning framework specifically designed for camera-controlled video re-rendering. Built upon a pretrained model, we optimize the model through a scale-aware perceptual reward mechanism. Specifically, we introduce a metric 3D estimator to extract precise camera trajectories from generated videos, explicitly penalizing deviations in rotation and translation. Furthermore, we meticulously designed a data pipeline strategy based on real-world conditioning videos and target camera trajectories derived from synthetic data, eliminating the reliance on paired data. Extensive experiments demonstrate that Geo-Align consistently outperforms existing supervised learning baselines in both precise camera controllability and visual fidelity, indicating the effectiveness of our method.