Make optimizer-updated parameters contiguous before optimizer creation.

Render one validation view and compute all requested image metrics. The train/test exposure mode compares only the held-out half of

Read an optimizer step counter whether it is stored as a tensor or int.

Read or generate the initial point cloud for a Colmap scene. `sfm` converts COLMAP points to PLY on first use. `random` creates the

Fill the Gaussian model either from a saved iteration or the initial point cloud. Signature checks keep compatibility with related G

Detect the dataset type from files on disk and load matching scene metadata.

Parse a metric for sorting, treating missing values as negative infinity.

Apply a keep mask to all optimizer parameters and matching state tensors.

Read Colmap camera extrinsics and intrinsics. Binary files are preferred because they are the normal COLMAP output; text fil

Read depth-scale metadata and attach the median valid scale to each entry.

Convert fields in `self._fields` into argparse options. A leading underscore means the option also gets a one-letter short

Replace `f_rest` and copy compatible optimizer momentum into the new tensor.

Convert a device string to torch.device, falling back to CUDA on failure.

Build the Colmap test-view name list from evaluation settings. LLFF-style scenes use every `llffhold`-th sorted image as test. Other

Resolve how many contiguous scalar values each Gaussian owns in one GaussianAdam parameter group.

Restore checkpoint optimizer state by parameter group name. Tensor states are copied only when their shape still matches the

Run ImprovedGS budgeted densification on densification intervals. Edge-aware scores are computed when enabled. The Gaussian model th

Apply the MCMC update by relocating weak Gaussians and adding new ones.

Apply MiniGS structure updates. Most intervals split blurred Gaussians with the MiniGS mask. At the reinitialization interva

Run the original 3DGS clone/split/prune rule at configured intervals.

Rebuild MiniGS Gaussians from depth back-sampled training pixels. Each view contributes samples from pixels with low accumulated alp

Pick the training views used for ImprovedGS edge-aware scoring. The function keeps a small rotating pool in runtime state so repeate

After an opacity reset, schedule RAP's delayed pruning step. The delayed step removes a configured low-opacity percentage after the

Choose the background color for this render. Random background is a training augmentation. Otherwise the dataset's fixed whi

Check whether the current iteration lands on an opacity-reset interval.

Check whether RAP should reset opacity on this iteration. RAP resets only inside the densification window and only on the normal

Apply the SSIM formula using prebuilt convolution windows.

Step either the custom GaussianAdam optimizer or the default Adam optimizer.

Ensure the model optimizer type matches the current config.

Copy the starting point cloud and camera metadata into the experiment folder.

Delegate MCMC Gaussian insertion and return the number of added points.

Keep only best-repeat outputs on disk and return rows for the winners. Non-winning repeat folders are removed. The winning folder is

Add GNS opacity regularization during the pruning window. The strength is adjusted every few iterations by comparing the current

Run GNS pruning logic for the current iteration. The first pruning iteration scales the opacity learning rate. During the pr

Remove Gaussians below the opacity threshold without crossing the budget. If threshold pruning would remove too many Gaussians, the

Add MCMC SGLD noise to positions through the method helper.

Add MCMC's fixed opacity and scale penalties to the current loss.

Select the regularization rule for the active training method.

Attach scene-dependent objects and initialize the runtime-state cache.

When one run contains multiple training jobs, back up the batch config into the output root.

Measure average FPS by repeatedly rendering all views. Warmup rounds let CUDA kernels and caches settle before the timed rounds.

Create the CUDA background tensor that matches the dataset setting.

Create a GaussianModel and attach method-specific initialization flags.

Build the text content for the standalone opacity report.

Build post-processing arguments from config and matching train settings. Dataset-related train options are copied so post-processing

Build a random initial point cloud for Colmap scenes. The random points are sampled inside a cube scaled by the scene radius,

Build the train/test split plan requested by CLI flags.

Build the context skeleton before scene creation. Method settings are parsed early and written back to `opt`, so later modul

Build shared loop state after scene creation. This attaches scene-related objects into `context.runtime_state`, creates the

Normalize the selected training method and resolve method component flags.

Build the training-parameter record saved in one experiment folder and add the start time.

Convert a Camera object into the compact JSON format saved with outputs.

Pack model tensors and optimizer state into the checkpoint tuple format.

Append new Gaussian tensors to optimizer parameters and extend state buffers.

Remove a GPU lock file when the process that created it is gone.

Read result rows from existing output folders without launching jobs.

Compute a grayscale edge map from a training image. The image is converted to an 8-bit grayscale approximation and filtered

Compute GNS pre-activation opacity regularization. GNS regularizes raw opacity logits instead of activated opacity values, w

Clamp sample counts and call the CUDA relocation kernel.

Build the ordered PLY attribute list for all saved Gaussian fields.

Check whether the current config contains batch training, which decides whether to back up the batch config.

Build a channel-wise 2D Gaussian window for SSIM convolution.

Run original 3DGS densification, then prune Gaussians that meet the pruning rules.

Select candidate Gaussians with absolute screen-space gradients, then run long-axis split under the budget. In late training, fall back t

Reuse the original clone branch, then merge the MiniGS blur mask during split to match the reference densification behavior.

Add the MiniGS blur mask to the original split candidates, so highly blurred areas can split even when gradients are weak.

Convert a payload dictionary into argparse-style `--key value` tokens.

Evaluate one split from start to finish. For each view, the function renders the image, optionally saves render and ground t

Return only scenes requested by the filter and report unknown names early.

After training ends, add the end time and pure training duration to the parameter record.

Flush the wrapped stdout stream.

Build a normalized 1D Gaussian kernel used by SSIM.

Build the perspective projection matrix used by the main renderer.

Return the average camera center and the largest distance from it.

Merge command-line arguments with the saved `cfg_args` file in model_path. Values provided on the command line take priority, while

Create method-specific runtime state after the scene and model exist. ImprovedGS/GNS precompute edge maps for edge-aware scoring. Mi

Register cleanup handlers for SIGINT and SIGTERM.

Convert opacity values from sigmoid space back to raw logits.

Check whether resume should reuse the current output folder. An empty checkpoint directory means the checkpoint is expected inside

Return mean absolute pixel error between prediction and ground truth.

Load one image/depth pair and convert a CameraInfo record into a Camera. Resolution handling follows the original 3DGS behavior: fix

Load one batch JSON config, returning an empty dict for missing content.

Load one `result_<split>.json` file if it exists and contains data.

Load checkpoint payload and validate that it matches the supported model type. The loader accepts both the current dictionary format

Sample candidates by importance, create two child Gaussians by moving along the longest axis in both directions, then adjust scale and op

Parse batch arguments and process every resolved config path.

Create a folder and ignore the error when it already exists.

Convert a parameter object into a plain dictionary that can be written to JSON.

Gate noise so low-opacity Gaussians receive stronger perturbations.