The base path that contains all of the inputs, outputs, and virtual environment used during the training process. Outputs are produced in specific subdirectories depending on the type of output file.

The name of the subdirectory that contains all outputs produced by this node.

The name of the training run. If this node is being wedged out to perform different training runs with different settings, this parameter can create a per-wedge output directory.

The path to the directory that contains all ML models, tests, and plots for the current training run. By default, this parameter is derived from the Base Directory, Base Name, and Run Name parameters, but can be modified to a custom directory path.

Parent directory containing the training data set as a named subfolder.

Name of the data set subdirectory inside Data set Folder.

The format of the input data set.

Integer factor used to downscale data set images before training. Higher values reduce memory usage and training time at the cost of detail.

Multi parameter list of auxiliary AOVs from the Houdini data set (e.g. normals or albedo) to include in training alongside RGB color. Each entry adds an extra per-Gaussian feature channel that is optimized against the corresponding AOV images.

AOV Name

The name of the AOV layer in the multi-layer EXR data set images.

Note

Most AOVs can be trained as extra features into a splat, but there are a few limitations to keep in mind

• Avoid view dependent AOVs (e.g. specular). The spherical harmonics are reserved for the RGB values.

• Avoid extremely high dynamic range (e.g. P depending on the scene scale). It makes sense to scale such values before training on them.

• Avoid contrast rich textures that display different details than the RGB. See more in Geometry Weight.

Signature

Number of channels in the feature: Mono (1), UV (2), or RGB (3).

Loss Weight

Weight for this feature in the loss relative to the RGB term.

Tip

Use the PDG ML Training Monitor to monitor the different loss magnitudes and fine tune the loss weight.

Geometry Weight

How strongly this feature’s gradient influences splat geometry (positions, scales, orients). A value of 0 trains the feature as appearance only, without affecting geometry. This means only the RGBA values of the target EXR image steer the optimization of the geometric components of the Gaussian splat. This leads to the best possible result for the overall appearance, but chokes the optimization of the extra features. Higher values can lead to artifacts in the RGB appearance of the Gaussian splats.

Tip

This setting depends highly on the actual data being fed to the training. It is always a trade-off between RGB quality and extra feature quality. The balance becomes increasingly more difficult with more extra features.

Note

Training for very contrast rich AOVs that contain texture information contrary to the RGB plane’s texture information can lead to artifacts. This happens because the training optimizes the Gaussian splats' appearance by moving around the separate splats to represent a given texture. When the RGB texture provides one alignment signal, but an additional feature such as a checkerboard texture provides a conflicting signal, the splats may align with the checkerboard edges. As a result, the checkerboard pattern can appear in the geometry of the entire model. In these cases, a fine balance between the Geometry Weight and other loss related hyperparameters is required. It also makes sense to aim for a higher point count to allow for more splats in areas that require those different detail levels.

When on, the RGB input is converted through OpenColorIO before being used as a training target. Extra features remain in their source colorspace.

Path to the OCIO configuration file used for the color transform.

The colorspace of the data set images.

The colorspace the model is trained against.

Directory where trained PLY results are written.

Filename prefix for exported gsplats files.

Turn on an additional USD export of the trained gsplats.

USD encoding: binary .usd or ASCII .usda.

When on, the scene is recentered and rescaled based on the training cameras before training. This improves numerical stability and keeps hyperparameters such as learning rates consistent across data sets of different physical sizes.

A global scaler that applies to all scene-size-related parameters.

How the initial Gaussian point cloud is seeded.

When on, a /pointcloud/points3d.ply file found in the data set folder is used as the initial point cloud instead of the SfM data. This allows feeding more accurate initialization data to the training that already includes more Gaussian splat properties beyond color and position in the regular SfM data set. The color is expected to be in the same colorspace as the source EXR images.

Path to the checkpoint PLY file to resume training from.

Step number to start counting from when resuming. This affects schedule- driven parameters such as learning rate decay and densification windows. It makes sense to set this to the previous training’s final step count.

Number of Gaussians created for random initialization.

Size of the cube, in world units, used to scatter random initial points.

Starting opacity for each Gaussian when using the Default strategy.

Starting scale for each Gaussian when using the Default strategy, expressed as a multiplier relative to the local point density.

Starting opacity for each Gaussian when using the MCMC strategy.

Starting scale for each Gaussian when using the MCMC strategy.

Random seed for parameter initialization (quaternions, scales, features, random points).

When on, adaptive densification and pruning are performed during training. When off, only the parameters of the initial Gaussians are optimized. No points are added or removed.

Densification algorithm.

Gaussians with opacity below this threshold are pruned (Default strategy).

Maximum number of Gaussians allowed during training (MCMC strategy). No new points will be added beyond this count. However, points that already existed during initialization will not be removed to reach that point count. To target a specific count the initial input cannot have more than the specified max value.

Opacity threshold below which Gaussians are considered dead and get relocated (MCMC strategy).

Amplitude of the positional noise injected by the MCMC strategy at each step (MCMC strategy).

Screen-space gradient threshold above which a Gaussian is considered under-reconstructed and densified (Default strategy).

World-space scale threshold separating duplication from splitting: Gaussians smaller than this are duplicated, larger are split (Default strategy).

Screen-space scale threshold used to also trigger splitting for Gaussians that project to large areas in the image (Default strategy).

Gaussians whose world-space scale exceeds this fraction of the scene extent are pruned (Default strategy).

Gaussians whose projected screen-space scale exceeds this fraction are pruned (Default strategy).

Use absolute screen-space gradients (AbsGS) instead of signed gradients when evaluating the growth criterion. Typically requires a larger Grow Grad 2D value (Default strategy).

Use the revised opacity heuristic when splitting/duplicating so the combined opacity of new Gaussians matches the original (Default strategy).

Blend factor between L1 and SSIM in the photometric loss. The total loss is (1 - lambda) * L1 + lambda * SSIM.

Controls how image alpha is used in the loss.

Weight of the direct alpha loss term.

Multiplier applied to the loss near alpha edges when alpha supervision is active. Higher values push the model to resolve silhouettes more accurately.

L1 regularization weight applied to Gaussian opacities (Default strategy).

L1 regularization weight applied to Gaussian scales (Default strategy).

L1 regularization weight applied to Gaussian opacities (MCMC strategy).

L1 regularization weight applied to Gaussian scales (MCMC strategy).

Random seed for the data set sampling order (image selection on each step).

Run test evaluations on held-out views during training.

Every Nth image in the data set is held out as a test image. The remaining images are used for training.

Total number of training steps.

Multiplier applied to all step-based schedules (densification windows, evaluation, checkpointing, PLY export). Useful for proportionally lengthening or shortening a run without re-tuning every individual interval.

Learning rate for Gaussian positions. This rate is scaled by the scene extent when Normalize World Space is on.

Learning rate for the diffuse RGB component (band 0 of the spherical harmonics coefficients).

Learning rate for higher-order spherical harmonics coefficients.

Learning rate for Gaussian opacities.

Learning rate for Gaussian scales.

Learning rate for Gaussian rotations.

Learning rate for extra features.

Every N steps, increase the active spherical harmonics degree by one up to the maximum. Lower degrees are trained first to stabilize color before higher frequency directional effects are introduced.

Step at which densification begins (Default strategy).

Step at which densification ends (Default strategy). The remaining steps only refine the existing Gaussians.

Every N steps, reset Gaussian opacities to a small value so that unused Gaussians can be pruned (Default strategy).

Perform a densification pass every N steps (Default strategy).

Step at which 2D-scale-based pruning stops. 0 disables this criterion (Default strategy).

Number of steps to skip densification after an opacity reset, allowing the optimizer to stabilize. The default value of 0 means no pause at all, but one might want to set this parameter to the number of images in the training set (total images in the data set - test images) to avoid pruning Gaussians before they could recover after a reset. This can occur with very low prune opacity thresholds (Default strategy).

Step at which densification begins (MCMC strategy).

Step at which densification ends (MCMC strategy).

Perform an MCMC refinement step every N steps.

Number of images processed per training step. Larger batches produce smoother gradients but use more GPU memory.

Random seed for per-step training randomness (random backgrounds, strategy split/noise).

Start the interactive gsplat web viewer during training so the current model can be inspected live in a web browser.

Open the viewer URL in the default web browser.

When on, automatically open the viewer URL in the default browser when training starts.

Host name the open viewer button targets. The host is always the machine running the training. This allows you to specify the IP of the training computer and open the viewer directly. This can be useful when running the training on a farm machine, but inspecting it from a work station.

TCP port the viewer attempts to serve on.

Save rendered validation images at each test step.

Encode a flythrough video from the test renders.

Directory for test outputs (metrics and rendered views).

Filename prefix for test files.

Run test evaluation every N training steps.

Echo log messages to standard output in addition to the log file.

Log detailed information for each densification operation.

Directory for training log files.

Filename prefix for log files.

Write a log entry every N training steps.

Directory for training plots.

Filename prefix for plot files.

Directory for PLY and USD checkpoint files.

Filename prefix for checkpoint files.

When on, export PLY checkpoint files at the interval specified by PLY Export Interval.

Export a PLY checkpoint every N steps.

When on, export USD checkpoint files at the interval specified by USD Export Interval.

Export a USD checkpoint every N steps.

Directory for training state checkpoints (model + optimizer state).

Filename prefix for state checkpoint files.

Determines how the processor node handles work items that report expected file results.

The path to the virtual environment on disk.

Determines whether the virtual environment should be created with symlinks to the source Python binaries. This is recommended on Linux and Mac when creating a virtual environment using Houdini’s Python interpreter.

When on, pip will attempt to use cached packages on the local system instead of downloading them every time. This can improve the installation times when repeatedly installing the same Python package in different virtual environments.

Keep decoded training images resident on the GPU. Significantly speeds up training at the cost of VRAM.

Convert data set EXRs to compressed NumPy .npz files once up front for faster subsequent loads. Only applies to Houdini data sets.

Strategy for choosing the number of data-loading worker processes.

Explicit worker count when Number of Workers is set to Custom Worker Count.

Pin training to a specific GPU instead of using the default device.

Index of the GPU to use when Override GPU is on.

Overrides the TOP scheduler for this node.

When on, specifies an expression that determines which work items from the node should be scheduled. If the expression returns zero for a given work item, that work item will immediately be marked as cooked instead of being queued with a scheduler. If the expression returns a non-zero value, the work item is scheduled normally.

Determines how the node should label its work items. This parameter allows you to assign non-unique label strings to your work items which are then used to identify the work items in the attribute panel, task bar, and scheduler job names.

When on, specifies a custom label for work items created by this node. The parameter can be an expression that includes references to work item attributes or built-in properties. For example, $OS: @pdg_frame will set the label of each work item based on its frame value.

Determines how the current scheduler prioritizes the work items in this node.

Specifies an expression for work item priority. The expression is evaluated for each work item in the node.

This parameter is only available when Work Item Priority is set to Custom Expression.