Summary
SCoPE is a camera-control method for video diffusion transformers. It gives every video token a camera ray as a second positional coordinate so target camera motion becomes part of the coordinate system.
SCoPE
NEWKey Features
Camera-ray positional encoding for video tokens.
Single-frame to target-trajectory generation.
Metric and up-to-scale pose support.
Lightweight Normalize-Gate-Inject module.
Preserves pretrained RoPE exactly.
Improved camera controllability.
Works at 5B and 14B scales.
Public code, model, and live demo links.
About SCoPE
Its lightweight Normalize-Gate-Inject path works with metric and up-to-scale poses while preserving pretrained rotary positional embeddings. Ray geometry inside attention reduces appearance ambiguity and helps the generated scene follow a commanded trajectory.
SCoPE is useful for controllable image-to-video generation, virtual cinematography, robotics visualization, and research into geometric conditioning. The project reports improved rotation error and FVD at both 5B and 14B model scales with fewer than 0.1% new parameters.
Get more likes & reach the top of search results by adding this button on your site!
Subscribe to the AI Search Newsletter
Get top updates in AI to your inbox every weekend. It's free!