Full camera record with image data and projection matrices.

r"""Creates a criterion that measures Learned Perceptual Image Patch Similarity (LPIPS). Arguments: net_type (str): the network type

Minimal point-cloud record with positions, RGB colors, and normals.

Lightweight camera metadata produced by dataset readers.

Simple object used to hold extracted parameters as attributes. The project historically reads options with dot access, so this class

COLMAP image record with a convenience quaternion-to-matrix method.

Dataset and output-path options used before the scene is created.

Renderer behavior options passed to the Gaussian rasterizer.

Loaded dataset plus the Gaussian model associated with it.

Complete dataset description consumed by Scene.

Stdout wrapper that appends timestamps unless quiet mode is enabled.

GNS opacity regularization and learning-rate scaling settings.

ImprovedGS structure-growth, budget, RAP, and MU schedule options.

MCMC relocation and noise regularization settings.

MiniGS-specific limits and periodic reinitialization settings.

Common optimization settings shared by all training methods.

Wrapper that registers and merges all optimization-related groups.

Final pruning budget settings used by regularized pruning methods.

Bundle long-lived objects needed by training stages. Passing this object keeps stage function signatures short while still m

Method selector and ImprovedGS component switches.

Trainable Gaussian collection used by rendering, training, and saving code.

Mixin that adds structure updates to GaussianModel.

Mixin that creates Gaussian tensors from point clouds and reads/writes PLY files.

Mixin that owns optimizer setup, restore, and tensor replacement logic.

Mixin that grows SH capacity and advances the active SH degree.

Lightweight camera view that stores only projection-related fields.

Base class that registers one group of argparse options from defaults.

Optional scratch fields shared across training stages. Values here are created once and updated over many iterations, such as

State reused by many training-loop stages. Keeping these values in one typed dictionary makes call sites shorter while still