Scope

SCoPE (Sightline-Coordinate Positional Encoding) is a research method from Tencent-affiliated authors for adding camera control to pretrained video diffusion transformers. Each video token receives its camera ray, derived from the user's camera trajectory, as a second positional coordinate. Ray geometry inside attention helps resolve appearance ambiguity and makes generation follow the target trajectory.

It is applied as a lightweight Normalize–Gate–Inject retrofit that supports both metric and up-to-scale poses and preserves the pretrained RoPE bit-exactly. It adds under 0.1% parameters and is reported to improve controllability and video fidelity at 5B and 14B scales. Creator coverage describes it as built on Wan 2.2 with code released under Apache 2.0.

The project page shows multiple trajectories per scene, WASD-style camera-driven motion of a third-person subject, and results on real captures, in-the-wild footage and stylized artwork. It also reports revisit consistency (returning to a previously seen region keeps it anchored) as an emergent property of the coordinate system rather than an added memory mechanism. It is a research framework rather than a hosted product; the page provided no repository or weights links.