SCoPE — steer the camera through a still image

SCoPE retrofits Wan2.2-I2V-A14B with Sightline-Coordinate Positional Encoding: Plücker camera rays are normalised, gated and injected straight into the DiT's self-attention queries/keys, so a real 3D camera path drives the generated shot.

Drop in an image, describe the scene, and pick a camera move — you get an 81-frame, 832x480, 16 fps clip that follows that trajectory.

Notes · Camera paths are OpenCV camera-to-world matrices [81, 3, 4] relative to the first frame, exactly the format SCoPE was trained on — the presets are taken from the release's own examples/ trajectory set. Non-16:9 uploads are centre-cropped.

To keep a run inside a ZeroGPU slot this Space serves both 14B experts in fp8 and samples with the Wan2.2-Lightning 4-step distillation LoRA at cfg_scale = 1.0 instead of the paper's 40 steps at cfg_scale = 3.5; expect slightly softer detail than the official samples.