Camera-control helpers for SANA-WM.
This module mirrors the public NVlabs/Sana SANA-WM camera preparation path: WASD/IJKL action rollout or explicit camera-to-world poses are converted into relative poses, per-latent-frame ray metadata, and per-VAE-chunk Plucker maps. Model-local projection layers consume these tensors in the Stage-1 DiT.
dataclass ¶
class-attribute instance-attribute ¶
pitch_limit_deg: float = SANA_WM_DEFAULT_PITCH_LIMIT_DEG
class-attribute instance-attribute ¶
rotation_speed_deg: float = (
SANA_WM_DEFAULT_ROTATION_SPEED_DEG
)
class-attribute instance-attribute ¶
translation_speed: float = SANA_WM_DEFAULT_TRANSLATION_SPEED
Frames action rolls out: the identity start pose plus one per step.
Totals the durations arithmetically, so an oversized request is rejected before anything allocates per-frame state.
action_string_to_c2w(
action: str,
*,
translation_speed: float = SANA_WM_DEFAULT_TRANSLATION_SPEED,
rotation_speed_deg: float = SANA_WM_DEFAULT_ROTATION_SPEED_DEG,
pitch_limit_deg: float = SANA_WM_DEFAULT_PITCH_LIMIT_DEG,
) -> ndarray
Roll out an OpenCV-convention (N + 1, 4, 4) c2w trajectory.
build_plucker_condition(
condition: SanaWmCameraCondition,
*,
vae_stride: tuple[int, int, int]
| list[int] = SANA_WM_DEFAULT_VAE_STRIDE,
) -> dict[str, Tensor]
Build native Stage-1 camera tensors from normalized request metadata.
compute_raymap(
intrinsics: Tensor,
poses: Tensor,
height: int,
width: int,
*,
use_plucker: bool = True,
) -> Tensor
Compute SANA-WM geometry ray maps.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
intrinsics | Tensor |
| required |
poses | Tensor |
| required |
height | int | raymap height in latent pixels. | required |
width | int | raymap width in latent pixels. | required |
use_plucker | bool | return | True |
Invert homogeneous rigid transforms with shape (..., 4, 4).
intrinsics_to_vec4_array(
intrinsics: Any,
*,
num_frames: int,
height: int,
width: int,
) -> ndarray
Normalize the intrinsics payload to an (F, 4) [fx, fy, cx, cy] array.
Accepts either None (derive from the output resolution) or the {fx, fy, cx, cy} mapping — the only form the request contract exposes.