Load pipeline, then try to attach `vggt` from subfolder unless explicitly provided.

Checks if 'input_tensor' contains inf or nan values and clamps extreme values. Args: input_tensor (torch.Tensor): The loss tenso

Sets the discrete timesteps used for the diffusion chain (to be run before inference). Args: num_inference_steps (`int`):

Convert a pose encoding back to camera extrinsics and intrinsics. This function performs the inverse operation of extri_intri_to_pose_encoding,

Calculate adaptive weight as 10^n so that weight * current_loss has the same order of magnitude as target_loss. Args: target_los

Returns the first found floating dtype in parameters if there is one, otherwise returns the last dtype it found.

Save pipeline as usual, then persist `vggt` into a subfolder if available.

Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with the multistep DPMSolver.

Unproject a batch of depth maps to 3D world coordinates. Args: depth_map (np.ndarray): Batch of depth maps of shape (S, H, W, 1) or

Implementation of the forward pass through the DPT head. This method processes a specific chunk of frames from the sequence.

Apply basic activation function to pose parameters. Args: pose_enc: Tensor containing encoded pose parameters act_type: Acti

Center crops a tensor of frames.

Clean the prompt to remove fixed prefixes.

Args: c2w (torch.tensor): [v, 4, 4] K (torch.tensor): [v, 3, 3] h (int): height of the image w (int): width of th

Apply inverse log transform: sign(y) * (exp(|y|) - 1) Args: y: Input tensor Returns: Transformed tensor

Load a clip from an extracted-frame directory, then apply "sampling/padding + aspect-preserving resize + center crop". Two frame-selecti

Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles custom timesteps. Any kwargs

Align predicted point cloud P to ground truth point cloud G using Umeyama algorithm. Args: P: (F, H, W, 3) predicted point cloud

Applies 1D rotary position embeddings along one dimension. Args: tokens: Input token features. positions: Position in

Apply positional embedding to tensor x.

"Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the prediction of x_0 at t

"Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the prediction of x_0 at t

Compute point-cloud reconstruction metrics after optional alignment and downsampling. This function aligns prediction `P` to ground truth `G

Compute depth loss. Args: gamma: Weight for confidence loss alpha: Weight for confidence regularization valid_ra

Compute point loss. Args: gamma: Weight for confidence loss alpha: Weight for confidence regularization valid_ra

Custom interpolate to avoid INT_MAX issues in nn.functional.interpolate.

Filter loss tensor by keeping only values below a certain quantile threshold. This helps remove outliers that could destabilize training

q: [B, Lq, Nq, C1]. k: [B, Lk, Nk, C1]. v: [B, Lk, Nk, C2]. Nq must be divisible by Nk. q_lens

this will perform the index select, cat the tensors, and provide the attn_bias from cache

This function generates a 1D positional embedding from a given grid using sine and cosine functions. Args: - embed_dim: The embedding di

Preprocess the poses to: 1. translate and rotate the scene so that the first frame is the identity

Core regression loss function with confidence weighting and optional gradient loss. Computes: 1. gamma * ||pred - gt||^2 * conf - al

Processes specialized tokens with shape (1, 2, X, C) for multi-frame processing: 1) Uses the first position (index=0) for the first frame onl

Build the patch embed layer. If 'conv', we use a simple PatchEmbed conv layer. Otherwise, we use a vision transformer.

Initializes the position generator with an empty cache.

Computes frequency components for rotary embeddings. Args: dim: Feature dimension (must be even). seq_len: Maximum se