WorldCast generates character video under large-angle camera motion
Original title: WorldCast: World-Consistent Character Video Generation with Large-Angle Camera Control
CLACT adds multi-view character data, and GAMA injects geometrically aligned features to limit identity drift.
WorldCast addresses character video generation when large combined camera motions reveal unseen regions and cause distortion and identity drift. The authors introduce CLACT, a character-centric dataset and benchmark with camera geometry and multi-view observations. The method back-projects the first frame into a depth-based point cloud, renders trajectory guidance, and uses Geometry-Aligned Memory Attention to retrieve view-relevant features. Evaluations report trajectory following while preserving geometry, appearance, and identity.
- CLACT covers diverse large-angle camera trajectories for training and evaluation.
- Point-cloud scale is aligned with the target camera trajectory.
- GAMA uses relative camera poses in a shared coordinate frame.
From the original article
1 min read at huggingface.co
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.
By Lei Cao, Yu Liu, Wenhua Liao · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.