WorldCast keeps characters coherent through large camera moves
WorldCast: World-Consistent Character Video Generation with Large-Angle Camera Control · arXiv:2610.33135
In plain words
When a video follows a camera moving a long way around a character, it can lose track of the path or change how the character looks.
WorldCast is a video-generation framework that uses the starting image to guide the camera movement and other character views to help fill in what the starting image cannot show.
On the authors’ CLACT evaluation, WorldCast scored best among the listed methods on camera-path errors and several character-consistency measures. Its visual-quality scores were competitive, but not best on every measure.
The research
When a camera moves far around a character, a video generator must follow the requested path while keeping the character consistent, including on sides missing from the starting image. How can one method address both demands as the viewpoint changes?
WorldCast combines rendered 3D guidance with aligned character views to follow large camera paths and preserve character consistency across viewpoints.
Can a generated character video follow a camera moving far from its starting view without drifting or changing the character? WorldCast combines guidance rendered from the starting image with character information retrieved from other viewpoints.
On CLACT, the authors compare methods using shared target paths and a common evaluation pipeline.
WorldCast scores best among listed methods on camera-path errors and several character-consistency measures, supporting the approach on these tests; its visual quality is competitive rather than uniformly best.
The evaluation covers 42 scenes and six prescribed trajectories per scene. It does not establish performance on more complex dynamic scenes or longer or interactive camera paths.
What the evidence shows
- A linked generation problem
WorldCast is a video-generation framework for following large camera paths while preserving a character’s shape, appearance, and identity.
Large viewpoint changes create a linked challenge: the video can drift from the requested camera path, while newly visible character regions are not constrained by the starting image.
- A test for changing views
The authors introduce CLACT, a character-focused dataset and evaluation benchmark for large-angle camera control and character consistency. Its evaluation uses 42 synthetic and real-world scenes, each with six prescribed camera trajectories.
This tests whether generated videos follow the requested movement and preserve character consistency as viewpoints change.
- What the comparison shows
On CLACT, the listed methods receive the same target trajectories, adapted to their input formats. A shared evaluation pipeline measures camera-path errors and character consistency across views.
WorldCast has the best reported scores among those methods for path errors, cross-view geometry, and identity or appearance similarities. Its visual-quality scores are competitive rather than best across all measures.
- Testing the memory branch
The authors also remove the entire GAMA branch from WorldCast. In that ablation, every reported camera-control, character-consistency, and visual-quality metric worsens relative to the full configuration.
This supports a role for the branch as a whole, though the test does not isolate its internal parts.
- A bounded training result
A separate ablation before GAMA adaptation finds that excluding the UE5 camera-motion collection worsens all three camera errors compared with full Stage 1.
The result supports the collection’s contribution to first-stage camera-control training, but does not establish its effect separately in the final system.
Deep
- Original paper authors
- Lei Cao, Yu Liu, Wenhua Liao, Peng Zhang, Fei Wang
- Submitted to arXiv
- License
- arXiv non-exclusive