SCoPE: Sightline-Coordinate Positional Encoding for Video Diffusion Transformers

Minghao Yin · Jiahao Lu · Wenbo Hu · Wang Zhao · Ying Shan · Kai Han

SCoPE adds camera sightlines as positional coordinates to a pretrained video diffusion transformer. Given a first frame, a text prompt, and a camera trajectory, it generates a video that follows the requested camera motion while preserving the original image-to-video prior. This repository is a self-contained release for Wan2.2-I2V-A14B: it contains everything required for inference, so a separate Wan2.2 checkpoint download is not needed.

Download

pip install -U huggingface_hub
hf download TencentARC/SCoPE --local-dir checkpoints/SCoPE

The checkpoint is approximately 67 GB. Keep both the checkpoint and the Hugging Face cache on local storage.

Usage

Install the SCoPE code. The released weights were trained and evaluated with PyTorch 2.9.1 (CUDA 12.8); because changing the PyTorch version can change the numerical output, we recommend reproducing this exact environment with uv:

git clone https://github.com/TencentARC/SCoPE.git
cd SCoPE
uv sync
source .venv/bin/activate

Generate a video with an example camera trajectory:

python inference.py \
  --model_path checkpoints/SCoPE \
  --case omni-misty-forest \
  --trajectory truck_right \
  --output_path outputs/omni-misty-forest.mp4

For custom inputs:

python inference.py \
  --model_path checkpoints/SCoPE \
  --input_image path/to/first_frame.png \
  --prompt "A person walks along a misty forest trail." \
  --camera_path path/to/camera_poses.npy \
  --x_fov 1.11847 \
  --output_path outputs/custom.mp4

Camera poses use OpenCV camera-to-world coordinates and must have shape [81, 3, 4] or [81, 4, 4]. x_fov is the horizontal field of view in radians; pinhole cameras use xi=0. See the GitHub repository for the full documentation, options, and demos.

Training data

SCoPE is trained with RealEstate10K, DL3DV, PanShot, and OmniWorld. The datasets use a common camera protocol: poses are expressed relative to the first camera and translation is normalized with per-clip near depth, while absolute scale is handled inside the model by a learned scale gate. Users are responsible for following the licenses and terms of the corresponding datasets.

Intended use and limitations

This model is intended for research on image-to-video generation and controllable camera motion. It inherits the visual capabilities, biases, safety limitations, and computational requirements of Wan2.2. Results may degrade for inaccurate camera poses or intrinsics, trajectories far outside the training distribution, large occlusions, or unusually fast camera motion.

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},
  year={2026}
}

Acknowledgements

SCoPE is built on Wan2.2 and DiffSynth-Studio. We thank the authors and contributors of these projects.

License

SCoPE is released under the Apache-2.0 License.