Parse command-line options for the batch runner.

Safely parse the legacy `cfg_args` Namespace(...) text format.

Convert a comma-separated scene filter into a set of scene names.

Build the training CLI parser, register all option groups, and parse argv.

Precompute one edge map per training camera for later scoring calls.

Prepare the experiment folder before training starts. Existing output is archived unless this run resumes in place. The current

Print one split's aggregate metrics in a compact readable block.

Return whether a local process id still exists.

Import and run the heavy post-processing implementation on demand.

Read GPU memory and utilization numbers from `nvidia-smi`.

Read Blender/NeRF transform JSON and convert frames into CameraInfo records.

see: src/base/reconstruction.cc void Reconstruction::ReadImagesBinary(const std::string& path) void Reconstruction::WriteImagesBinary(const s

Taken from https://github.com/colmap/colmap/blob/dev/scripts/python/read_write_model.py

see: src/base/reconstruction.cc void Reconstruction::WriteCamerasBinary(const std::string& path) void Reconstruction::ReadCamerasBinary(const

Taken from https://github.com/colmap/colmap/blob/dev/scripts/python/read_write_model.py

see: src/base/reconstruction.cc void Reconstruction::ReadPoints3DBinary(const std::string& path) void Reconstruction::WritePoints3DBinary(con

see: src/base/reconstruction.cc void Reconstruction::ReadPoints3DText(const std::string& path) void Reconstruction::WritePoints3DText(const s

Read pure training time saved by the training script. A legacy metadata file is checked as a fallback so old outputs can still

Rebuild the Gaussian set from MiniGS periodic back-sampling results. Keep the current SH settings and exposure parameters so `training_se

Remove a lock file after the scene job finishes.

Delegate MCMC relocation of low-opacity Gaussians to the method helper.

Render one camera view with the standard 3DGS parameter tensors. The function always passes DC color, higher-order SH, scale, rotati

Render MiniGS auxiliary statistics for the current view. The CUDA path returns projected area and accumulated weights. MiniGS uses

Render MiniGS depth back-sampling data for one view. The returned 3D points, alpha map, and Gaussian indices are used when M

Return the output folder for one split and model iteration.

Replace one optimizer parameter tensor and reset its momentum buffers.

Resolve either a single config file or all runnable configs in a folder.

Choose the starting SH degree for the Gaussian model. MiniGS starts from degree 0 and restores higher SH capacity later; other

Return the checkpoint path to load, or None when resume is disabled. If the checkpoint directory is omitted, the current model outpu

Restore model tensors, gradient accumulators, and optimizer state.

Dispatch one iteration's densification work for the active method. The caller already produced `render_state` during optimization. T

Render one camera and create gradients for this iteration. The returned `render_state` keeps only the render outputs needed by later

Apply optimizer updates for one iteration. Exposure is always stepped with its own Adam optimizer. Dense Gaussian tensors fo

Dispatch RAP and GNS pruning actions for one training iteration.

Run scheduled evaluation, model saving, and checkpoint saving. The return value is the time spent in reporting, which the caller

Execute or rebuild one batch config. The function owns the top-level order for one config: validate scenes, prepare output r

Draw one random training camera without replacement. When the local pool becomes empty, it is refilled from the scene so each

Write current raw Gaussian tensors to a PLY point-cloud file.

Save Gaussian model state and iteration in the current checkpoint format.

Return the largest saved iteration number from an iteration folder.

Choose the winning repeat output name for each scene by PSNR. Test rows are preferred when available; train rows are used as a fallb

Return the GPU id to expose to the child process plus its lock path.

Register activation and inverse-activation functions for raw parameters.

Decide whether this iteration should step trainable parameters. ImprovedGS/GNS can use MU, which skips dense optimizer steps later i

Count Gaussians in opacity intervals to make model sparsity easy to inspect.

Run one complete training job and return pure training time in seconds. The function keeps the high-level training order in one plac

Evaluate train and test cameras and write a compact report. The function averages metrics over each split, prints them through tqdm,

Update learning rate and SH degree before rendering the current iteration.

Update exposure and xyz learning rates for the current iteration.

Update the tqdm postfix with smoothed loss and current Gaussian count.

Save opacity interval statistics next to the experiment output.

Write the parameters used by this run into the experiment folder for reproduction and post-processing.

Create an empty Gaussian model with the requested SH capacity.

Load cameras, scene bounds, and Gaussian state for one experiment.

Load image/depth tensors and precompute transforms for rendering.

Store projection fields and derive the camera center from the view matrix.

Add new Gaussians by sampling existing ones until the budget cap is reached.

Add covariance-shaped position noise scaled by opacity and xyz learning rate.

Extract model options and normalize the source path to an absolute path.

Merge all optimization parameter groups into one object for training.

Return DC color SH coefficients.

Return higher-order SH color coefficients.

Return activated opacity values in `[0, 1]`.

Return normalized quaternion rotations.