Tencent launches SCoPE: camera motion control in video diffusion without retraining

Researchers at TencentARC released a model that adds line-of-sight camera coordinates as a positional encoding to the existing Wan2.2-I2V-A14B model, allowing precise reconstruction of a camera trajectory from a single initial frame and a text prompt without modifying the base weights. The full checkpoint is approximately 67 gigabytes and contains everything required for inference; a separate download of Wan2.2 is not needed.

SCoPE accepts three inputs: the first frame, a textual description, and a camera path in OpenCV camera-to-world format with shape [81, 3, 4] or [81, 4, 4]. The parameter x_fov specifies the horizontal field of view in radians (pinhole cameras use xi = 0). The model was trained on four datasets—RealEstate10K, DL3DV, PanShot and OmniWorld—all adhering to a shared camera protocol: poses expressed relative to the first frame, normalized translation based on the nearest clip depth, and an absolute scale handled internally by a learned scale gate.

The weights are released under Apache-2.0 and were validated with PyTorch 2.9.1 on CUDA 12.8; the developers advise reproducing the exact environment with uv because altering the PyTorch version changes numeric output. Documentation notes that performance may degrade for inaccurate poses, trajectories far from the training distribution, large elevations, or especially rapid camera motion. The model inherits the visual capabilities, hallucination tendencies, bias constraints and computational demands of Wan2.2.

Built on Wan2.2 and DiffSynth-Studio, the code is available on GitHub with complete documentation, execution options and demos. It is a research release aimed at video generation from a single image with controllable camera motion, not a polished production solution. Users intending to run it will need local storage for both the checkpoint and the Hugging Face cache, as well as a GPU with sufficient VRAM to accommodate a 14-billion-parameter model.