SCoPE: Sightline-Coordinate Positional Encoding for Video Diffusion Transformers

Minghao Yin1* Jiahao Lu2* Wenbo Hu3† Wang Zhao3 Ying Shan3 Kai Han1‡
1The University of Hong Kong 2The Hong Kong University of Science and Technology 3ARC Lab, Tencent

* Equal contribution  ·  † Project lead  ·  ‡ Corresponding author

Each clip is generated from a single input frame and a target camera trajectory; the inset shows the camera motion. Use the arrows to browse different controls of the same scene.

Abstract

Video diffusion transformers address their tokens by position on the pixel-time grid: an address in the tensor, not in the world. The address we would want, the world point a token depicts, lies on a surface not yet generated, while its camera ray is fixed once the user specifies a trajectory. SCoPE therefore treats the ray as a second positional coordinate, and camera control becomes a property of the coordinate system, not an added module. The ray is added to the pretrained attention's queries and keys, and the score gains a term that reads the two rays alone. Its canonical form, the reciprocal product of line geometry, measures how nearly two lines of sight meet. Normalize-Gate-Inject makes a single encoding trainable across metric and up-to-scale pose sources. The retrofit keeps RoPE bit-exact, starts from the unchanged pretrained DiT, and adds under 0.1% new parameters. On Wan2.2 at 5B and 14B under matched data and budget, SCoPE improves every camera-controllability and fidelity metric, and the margins widen with model size. At 14B, rotation error falls 29% and FVD 43% below the strongest baseline on each metric.

SCoPE gives each token its camera ray as a second coordinate, using ray geometry inside attention to resolve appearance ambiguity and follow a target camera trajectory.
Camera control as a property of the coordinate system. SCoPE gives every token a second address, its camera ray (left). Rays are fixed by the user's trajectory while the scene, dashed, is not yet generated (center). Appearance cannot choose between the two identical windows in frame t+1. Only one of their rays meets the ray from frame t. Added to the pretrained queries and keys, the ray gives the score exactly this test, zero when rays meet, growing with the miss (right). Content identifies, cross terms arbitrate, geometry filters.

Key Ideas


Same scene · many cameras

Scenes

One generated scene browsed under six different camera trajectories. The inset visualizes the commanded camera path.

WASD · third-person control

Motions

WASD inputs drive the camera trajectory that moves the third-person subject. Overlay keys reflect the per-frame motion.

More results

Real-World Scenes

Diverse real captures driven along varied camera paths from a single frame.

More results

In-the-Wild Videos

Everyday footage re-rendered under controllable camera motion.

More results

Stylized & Artistic Scenes

Illustrated and painterly inputs animated with 3D-consistent camera control.

Citation

@article{yin2026scope,
  title   = {SCoPE: Sightline-Coordinate Positional Encoding for Video Diffusion Transformers},
  author  = {Yin, Minghao and Lu, Jiahao and Hu, Wenbo and Zhao, Wang and Shan, Ying and Han, Kai},
  journal = {arXiv preprint arXiv:2606.27345},
  year    = {2026}
}