save cfgs together to enable easily loading the pretrained model

A history segment that contains a frequency range that changes linearly over time.

Check if points are inside camera frustums. Args: points: Points to test, shape [B, P, Q, 3] where:

Only a single clip per video.

Given a nested list or dict return the desired value at key expanding callable nodes if necessary and :attr:`expand` is ``True``. The expansion

Retrieve support frames based on a distance function between conditions Args: query: Query conditions with shape [B, h, 16] - query

Saves an EMA copy of the model + EMA optimizer states for resume.

stable diffusion's noise schedule https://github.com/CompVis/stable-diffusion/blob/21f890f9da3cfbeaba8e2ac3c425ee9e998d5229/ldm/modules/diffu

Loss function defined over sequence of flow predictions

sigmoid schedule proposed in https://arxiv.org/abs/2212.11972 - Figure 8 better for images > 64x64, when used during training

History guidance scheme equivalent to: conditional sampling with stabilization technique (Chen et al., https://arxiv.org/abs/2407.013

History guidance scheme equivalent to: conditional sampling with stabilization technique (Chen et al., https://arxiv.org/abs/2407.013

History guidance scheme equivalent to: HG-f with stabilization

History guidance scheme equivalent to: HG-v with stabilization

r""" A context manager to in-place swap regular parameters with EMA parameters. It swaps back to the original regular paramete

History guidance scheme equivalent to: Temporal History Guidance (HG-t)

All testing happens here

All training happens here

r"""Here you compute and return the training loss and some additional metrics for e.g. the progress bar or logger. Args:

Update the state with extracted features. Args: features: Features of shape (B, num_classes). The features are logits ext

Update the metric with predicted videos. Args: preds: Predicted videos of shape (B, T, C, H, W), range [0, 1]

Update internal states with lpips score.

Check if using deterministic subdataset for evaluation

Check if subdataset strategy is enabled

All validation happens here

History guidance scheme equivalent to: Vanilla History Guidance (HG-v)

Convert versions of the form vX to X. For example, v12 to 12.

Visualize the conditions (camera poses) for a video. Args: conditions: Tensor of shape [B, T, 16] containing camera poses