Skip to content

vllm_omni.diffusion.models.lingbot_world.camera

Camera trajectory loading and geometry for LingBot World v2 conditioning.

CameraTrajectory dataclass

Camera-to-world poses and (fx, fy, cx, cy) intrinsics by frame.

intrinsics instance-attribute 

intrinsics: Tensor

poses instance-attribute 

poses: Tensor

TrustedActionDirectory dataclass

Canonical action location tied to the trusted root's filesystem identity.

relative instance-attribute 

relative: Path

root instance-attribute 

root: Path

root_device instance-attribute 

root_device: int

root_inode instance-attribute 

root_inode: int

build_plucker_embedding 

build_plucker_embedding(
    trajectory: CameraTrajectory,
    *,
    height: int,
    width: int,
    target_height: int,
    target_width: int,
    device: device,
    dtype: dtype,
    translation_scale: float | None = None,
) -> Tensor

Build checkpoint-ordered (ray origin, ray direction) channels.

interpolate_camera_trajectory 

interpolate_camera_trajectory(
    trajectory: CameraTrajectory, num_frames: int
) -> CameraTrajectory

Resample translations/intrinsics linearly and rotations with quaternion SLERP.

load_camera_trajectory 

load_camera_trajectory(
    action_directory: TrustedActionDirectory,
) -> CameraTrajectory

Load bounded numeric camera arrays through a trusted directory handle.

resolve_trusted_action_directory 

resolve_trusted_action_directory(
    action_path: str | PathLike[str],
    trusted_root: str | PathLike[str],
) -> TrustedActionDirectory

Resolve an action directory beneath a trusted root without exposing paths.