TencentARC Releases SCoPE — Camera Control in Video Diffusion Models via a Coordinate System

🛠️ SCoPE: Camera Control via Coordinates

TencentARC has released SCoPE, a camera control method for video diffusion generators using a coordinate system. Each video token receives a camera ray as a second position, allowing the model to understand 3D geometry. It reduces rotation error by 29% and FVD by 43%, with less than 0.1% additional weights. Apache-2.0.

🌍 The 'camera as coordinates' architecture is a generalizable pattern for video transformers. It works on an unchanged backbone.

👤 Download it, set a camera trajectory, and get a video with controlled motion. Python, CUDA-enabled GPU.

Source 1: https://arxiv.org/abs/2606.27345 Source 2: https://visual-ai.github.io/scope/