Interpolate log-linearly from `v0` (t=0) to `v1` (t=1).

Compute MSE given a PSNR (we assume the maximum pixel value is 1).

Volumetric rendering. Args: ray_batch: array of shape [batch_size, ...]. All information necessary for sampling along a ray, includi

Prepares inputs and applies network 'fn'.

Transforms input poses to lie roughly on the upper unit hemisphere.

Remove the homogeneous bottom row from [..., 4, 4] pose matrices.

The mip-NeRF Model. Args: rng: random number generator (or None for deterministic output). rays: util.Rays, a pytree of ray origins,

Evaluate the MLP. Args: rng: random number generator (or None for deterministic output). samples: a tuple containing: - mean:

Get the next training batch or test example. Returns: batch: dict, has 'rgb' and 'rays'.

Generating downsampled images.

Generating downsampled images.

Generates random full image rays for DietNeRF loss.

Generating random rays for full images.

Generates random rays.

Generating rays for all images.

Load images from disk.

Load images from disk.

Sample next test example.

Sample next training batch.

Initialize training.

Helper function for computing alpha compositing weights.

The 'average' error used in the paper.

Computes an error between two images using a (ternary) census transform. This error is large when things are very wrong, and will be 0 when the mat

Returns TV norm for input values. Note: The weighting / masking term was necessary to avoid degenerate solutions on GPU; only observed on individ

Computes loss weight for tv norm.

Approximate a conical frustum as a Gaussian distribution (mean+cov). Assumes the ray is originating from the origin, and base_radius is the radiu

Construct a Neural Radiance Field. Args: rng: jnp.ndarray. Random number generator. rays: an example of input Rays. config: A Config cl

Construct a bijection between metric distances and normalized distances. See the text around Equation 11 in https://arxiv.org/abs/2111.12077 for a

Creates lookat matrix.

Approximate a cylinder as a Gaussian distribution (mean+cov). Assumes the ray is originating from the origin, and radius is the radius. Does not

Behaves like jax.tree_map but doesn't recurse on fields of a dataclass.

A generalization of linspace(), geomspace(), and NeRF's "lindisp". Behaves like jnp.linspace(), except it allows an optional function that "curve

Returns a frequency mask for position encoding in NeRF. Args: pos_enc_length (int): Length of the position encoding. current_iter (int

Encode `x` with sinusoids scaled by 2^[min_deg:max_deg-1]. Args: x_coord: a tuple containing: x, jnp.ndarray, variables to be encoded. Should

Continuous learning rate decay function. The returned rate is lr_init when step=0 and lr_final when step=max_steps, and is log-linearly interpola

Load a CLIP model Parameters ---------- name : str A model name listed by `clip.available_models()`, or the path to a model check

Loads config.

Loads a split of a dataset using the data_loader specified by `config`.

Load EXIF data for an image.

Load an image and cast to float32.

Compute iint w[i] w[j] |t[i] - t[j]| di dj.

Computes the occulusion regularization loss. Args: rgb (jnp.array): The RGB rays/images. density (jnp.array): The current de