scale alphas_cumprod so that SNR is multiplied by shift ** 2

Test that identical cameras have FOV distance of 0.

Test distance metrics on a panorama trajectory

Convert the history segment to two partial history conditions, whose conditional scores are subtracted to get the composed score.

Returns the rays represented as positional encoding. Follows NeRF to map the rays into a higher-dimensional space. Args:

Training step

Unwrap shortcuts by replacing them with commands from corresponding yaml files. All shortcuts should be in the form of `@shortcut_name`.

Update the state with the video. Args: videos: Videos of shape (B, T, C, H, W) and range [0, 1].

Upsample flow field [H/8, W/8, 2] -> [H, W, 2] using convex combination

Wrapper that enables a function that operates on a batch of images to operate on a batch of videos. Can also be used as a decorator.

Weight the score (shape: (,)) with the dimension weight.

Compute distances between query and candidate conditions Args: query_conditions: Tensor of shape (B, h, C) c

Compute rotation angle differences between query and candidate poses Args: query_conditions: Query camera poses with sha

Compute L2 distances between query and candidate camera positions Args: query_conditions: Query camera poses with shape

Compute L2 distances between query and candidate image centers Args: query_conditions: Query camera poses with shape [B,

Compute weighted combination of rotation and translation distances Args: query_conditions: Query camera poses with shape

Compute distances that decrease linearly with time, ignoring actual poses. First frame gets minimal distance (maximal similarity), in

Compute distances based on field of view overlap between cameras. Smaller distance means more overlap. Args: que

Initialize the history guidance manager. This should be called at every sampling step. Args ---- mask: torch

Precompute all partial history conditions and corresponding weights needed for sampling. While history guidance can be computed by do

loads video together with the path where latent should be saved

Args: origin (torch.Tensor): The origin of the rays. Shape (B, T, H, W, 3). direction (torch.Tensor): The direction o

Args: R (torch.Tensor): The rotation matrix. Shape (B, T, 3, 3). T (torch.Tensor): The translation vector. Shape (B,