WorldCast: World-Consistent Character Video Generation with Large-Angle Camera Control

01 · Abstract

Camera moves. The world holds.

Precise camera control is essential for cinematic character video generation, yet existing methods struggle to follow large-angle trajectories while preserving character identity across viewpoint changes. We introduce CLACT, a character-centric benchmark of 80,000 training samples and 42 evaluation scenes, and WorldCast, a framework that couples scale-aligned geometric guidance with geometry-aligned multi-view character memory (GAMA). WorldCast achieves the best camera-control accuracy and character consistency while maintaining competitive visual quality.

WorldCast teaser: generated sequences along tilted circular and lemniscate trajectories, and quantitative comparison on camera accuracy, consistency, and quality

Figure 1 · Large-angle trajectories that break prior methods — best camera accuracy (ATE ↓), cross-view consistency (MEt3R ↓), and identity preservation, at competitive image quality.

DATASET

COMPONENT 01

CLACT Benchmark

A character-centric dataset & benchmark for Large-Angle Camera Trajectories: 80,000 synthetic training samples (50K UE5 + 30K Blender), 42 synthetic and real-world evaluation scenes, accurate camera geometry, and paired multi-view character observations. Dataset · coming soon

GUIDANCE

COMPONENT 02

Scale-Aligned Geometric Guidance

The initial view is lifted to a colored 3D point cloud with calibrated depth scale, then splat-rendered along the target trajectory — producing trajectory-aligned RGB guidance and coverage masks that tell the model exactly what to complete.

MEMORY

COMPONENT 03

Geometry-Aligned Memory Attention

Multi-view anchor geometry is aligned with the target trajectory in a shared coordinate frame via Umeyama alignment, enabling geometry-aware retrieval and injection of view-relevant character features into video tokens.

03 · Method

From one frame to a world model

WorldCast couples scale-aligned geometric guidance with geometry-aligned multi-view character memory, built on the Wan2.1-I2V-14B flow-matching backbone.

Figure 2 · WorldCast overview — trajectory-aligned geometric guidance coupled with geometry-aligned multi-view character memory (GAMA).

04 · Results

State of the art on CLACT

Evaluated on 42 character-centric scenes × 6 trajectories. WorldCast leads on every camera-control and character-consistency metric while keeping competitive visual quality.

Method	Anchors	ATE ↓	RTE ↓	RRE ↓	MEt3R ↓	ArcFace ↑	CLIP ↑	DINO ↑	Subj. ↑	IQ ↑
Camera-Controlled Generation
CamI2V	✗	0.351	0.281	11.490	0.405	0.147	0.804	0.675	0.874	0.686
DualCamCtrl	✗	0.297	0.082	3.607	0.408	0.185	0.796	0.661	0.904	0.699
GEN3C	✗	0.130	0.038	1.255	0.306	0.433	0.819	0.696	0.915	0.716
Uni3C	✗	0.098	0.033	1.069	0.324	0.374	0.823	0.691	0.911	0.711
Lyra 2.0	✗	0.142	0.034	1.741	0.316	0.394	0.817	0.691	0.916	0.725
Novel-View Generation
ViewCrafter	✗	0.353	0.203	9.198	0.395	0.223	0.810	0.683	0.888	0.715
ViewCrafter ⋆	✓	0.361	0.196	8.902	0.419	0.195	0.795	0.655	0.869	0.715
MVGenMaster	✗	0.125	0.099	2.929	0.370	0.227	0.796	0.655	0.869	0.699
MVGenMaster ⋆	✓	0.218	0.152	6.938	0.301	0.299	0.821	0.707	0.862	0.691
Video Re-Rendering
ReCamMaster	✗	0.295	0.067	2.600	0.335	0.363	0.812	0.678	0.904	0.689
DepthDirector	✗	0.289	0.066	1.732	0.332	0.423	0.826	0.686	0.919	0.705
CameraAnything	✗	0.244	0.078	4.879	0.341	0.358	0.807	0.648	0.898	0.700
Ours
WorldCast	✓	0.065	0.020	0.896	0.228	0.555	0.830	0.719	0.917	0.719

White = best per column · Underlined = second best · ⋆ = additional character views provided as input

Camera Accuracy

ABSOLUTE TRAJECTORY ERROR · LOWER IS BETTER

WorldCast0.065

Uni3C0.098

MVGenMaster0.125

GEN3C0.130

Lyra 2.00.142

Geometric Consistency

MET3R · LOWER IS BETTER

WorldCast0.228

MVGenMaster ⋆0.301

GEN3C0.306

Lyra 2.00.316

Uni3C0.324

Identity Preservation

ARCFACE SIMILARITY · HIGHER IS BETTER

WorldCast0.555

GEN3C0.433

DepthDirector0.423

Lyra 2.00.394

Uni3C0.374

05 · Comparison

The arena

Large-angle orbit with push-pull — the hardest shot in the benchmark. Same input, same trajectory, thirteen methods. Watch the character survive.

case 01 · orbit 360° + push-pull

case 02 · orbit 360° + push-pull

Input & Anchors

Camera Trajectory

WorldCast · Ours

Uni3C

GEN3C

Lyra 2.0

MVGenMaster +anc

MVGenMaster

ViewCrafter +anc

ViewCrafter

ReCamMaster

DepthDirector

DualCamCtrl

CameraAnything

07 · Citation

Cite WorldCast

@misc{worldcast,
  title         = {WorldCast: World-Consistent Character Video Generation
                 with Large-Angle Camera Control},
  author        = {Cao, Lei and Liu, Yu and Liao, Wenhua and Zhang, Peng and
                 Wang, Fei and Zhang, Hao and Le, Jiajun and Zhou, Daiguo and
                 Luan, Jian and Ma, Jiayi},
  year          = {2026},
  eprint        = {2610.33135},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2610.33135}
}

WorldCast