Forward pass of the ResNet block. Args: x: Input tensor of shape (B, C, H, W). emb: Embedding tensor of shape

Forward pass of the conditioning block. Args: x: Input tensor of shape (B, N, C). emb: Embedding tensor of sh

Forward pass of the attention block. Args: x: Input tensor of shape (B, N, C). emb: Embedding tensor of shape

Forward pass of the transformer block. Args: x: Input tensor of shape (B, N, C). emb: Embedding tensor of sha

Forward pass of the U-ViT backbone. Args: x: Input tensor of shape (B, T, C, H, W). noise_levels: Noise level

Forward pass of the DiTBase model. Args: x: Input tensor of shape (B, N, C). c: Conditioning tensor of shape

Forward pass of the AdaLN layer. Args: x: Input tensor of shape (B, N, C). c: Conditioning tensor of shape (B

Forward pass of the AdaLN-Zero layer. Args: x: Input tensor of shape (B, N, C). c: Conditioning tensor of sha

Forward pass of the DiT block. In original implementation, conditioning is uniform across all tokens in the sequence. Here, we extend

Forward pass of the DiT final layer. Args: x: Input tensor of shape (B, N, C). c: Conditioning tensor of shap

x ~ (b t c) attn_mask ~ (b t t)

History guidance scheme equivalent to: Fractional History Guidance (HG-f)

Get the frustum of the camera. Args: K: [..., 4] or [..., 3, 3] uv_range: [H, W] or [[H_min, H_max], [W_min, W_max]]

Only filter videos with camera rotation within the specified range.

Log gradient statistics such as the mean or std of norm.

Log image(s) using WandbLogger. Args: key: the name of the video. image: a single image or a batch of images.

Log video to wandb. WandbLogger in pytorch lightning does not support video logging yet, so we call wandb directly. Args:

Return the maximum logSNR value.

Return the minimum logSNR value.

Number of Function Evaluations (NFE) per batch & sampling step.

# assuming mask.shape = (batch_size, nfe, seq_len) return self.mask.shape[1]

Normalizes the camera poses by the first camera, i.e. computes the relative poses w.r.t. the first camera. After normalization, the f

Preprocess the batch before training/validation. Args: batch (Dict): The batch of data. Contains "videos" or "latents",

Enable loading EMA-enabled optimizer states to EMA-disabled optimizer states.