Volinga Suite provides multiple training strategies for reconstructing 3D Gaussian Splatting (3DGS) assets.
Although all strategies optimize a set of 3D Gaussian primitives to reproduce the input views, they differ significantly in how the Gaussian population evolves during training.
A training strategy determines how the system identifies under-reconstructed regions, how new Gaussians are introduced or existing ones are redistributed, how redundant primitives are removed, and how the available Gaussian budget is allocated across the scene.
These decisions directly affect:
Volinga Suite provides the following training strategies:
The strategies available in the Trainer depend on the installed version of Volinga Suite. Refer to the Training Parameters section in Getting Started for the corresponding version compatibility table.
MCMC uses a fundamentally different approach to Gaussian population management than traditional 3DGS densification.
Instead of treating Gaussians as a set of primitives that must be explicitly cloned, split, and pruned according to a collection of geometric thresholds, MCMC interprets the Gaussian population as a set of samples drawn from an underlying probability distribution representing the scene.
During optimization, Gaussian positions are updated using a formulation based on Stochastic Gradient Langevin Dynamics (SGLD). In addition to the gradient produced by the reconstruction objective, controlled stochastic noise is injected into the positional updates.
This stochastic component allows Gaussians to continue exploring the reconstruction space rather than following only deterministic gradient descent trajectories.
As training progresses, Gaussians that contribute little to the reconstruction can be relocated toward regions where additional representation capacity is required. This effectively transforms conventional densification and pruning into a redistribution process.
The result is a highly exploratory optimization strategy capable of recovering scene regions that may have been poorly represented by the initial point distribution.
MCMC operates with a predefined maximum Gaussian budget.
Rather than allowing the number of primitives to grow indefinitely through repeated cloning and splitting, the strategy progressively makes use of the available budget and redistributes Gaussian capacity throughout the scene.
This makes the Max Gaussians parameter particularly important when using MCMC, as it establishes the upper bound of the representation.
The stochastic nature of the optimization also means that two training runs using identical source data and parameters may not produce exactly the same Gaussian distribution.
Strong global exploration
Stochastic positional updates allow Gaussians to explore regions that deterministic optimization may fail to reach.
Robustness to imperfect initialization
The method is less dependent on having an ideal initial COLMAP point cloud because primitives can migrate toward under-represented regions during optimization.
Effective recovery of poorly covered regions
Sparse or initially missing areas can receive representation capacity as low-contribution Gaussians are redistributed.
Explicit Gaussian budget control
The maximum primitive count can be defined in advance, making memory and asset complexity easier to constrain.
Reduced dependence on handcrafted densification rules
Gaussian redistribution replaces much of the conventional clone/split decision logic.
Non-deterministic optimization
The stochastic component means that repeated training runs can converge to slightly different representations.
Potentially less structured Gaussian distributions
Because Gaussians continuously explore and relocate, the final distribution can be less spatially regular than strategies based on targeted geometric splitting.
Fine edges may appear softer
MCMC prioritizes robust scene coverage and exploration rather than explicitly targeting high-frequency image structures during densification.
Convergence may require additional iterations
Exploration can remain active for longer before the Gaussian distribution stabilizes.
Gaussian budget selection matters
A budget that is too small can limit reconstruction capacity, while an unnecessarily large budget increases memory consumption and final asset size.
MCMC is particularly suitable for:
When to choose MCMC
Choose MCMC when the input dataset is difficult, the initialization is not ideal, or when other strategies fail to adequately cover parts of the scene.
MRNF is a hybrid refinement strategy designed to combine the exploration capabilities of stochastic optimization with the spatial precision of targeted Gaussian splitting.
Instead of relying exclusively on stochastic Gaussian redistribution or exclusively on deterministic densification thresholds, MRNF combines multiple refinement signals to determine where additional representation capacity is required and how that capacity should be introduced.
The strategy combines:
This allows MRNF to simultaneously address two important problems in Gaussian reconstruction:
MRNF analyzes reconstruction information in image space to identify areas where the current Gaussian representation does not sufficiently reproduce the training images.
Instead of uniformly increasing Gaussian density, refinement can therefore be concentrated in regions associated with reconstruction error.
This helps prevent the Gaussian population from growing equally across both simple and complex regions of the scene.
High-frequency image structures such as object silhouettes, thin geometry, texture boundaries, and sharp transitions often require a denser or more accurately positioned Gaussian representation.
MRNF incorporates edge information into the refinement process so that these regions can receive additional representation capacity.
This generally improves the reconstruction of:
However, edge information does not always correspond to geometric detail. High-frequency image noise, foliage, reflections, or strongly textured surfaces may also generate significant edge responses.
For this reason, edge-guided refinement must be balanced against the other signals used by the strategy.
When refinement is required, MRNF can use the IGS+ long-axis splitting strategy.
A 3D Gaussian is anisotropic: its covariance defines an oriented ellipsoid with different spatial scales along its principal axes.
Instead of generating children through generic spatial sampling, long-axis splitting uses the geometry of the Gaussian itself and divides it along its dominant axis.
This produces child Gaussians that more closely follow the spatial structure represented by the parent primitive and can reduce unnecessary overlap between newly created Gaussians.
The approach is particularly effective around elongated surfaces, boundaries, and thin structures where conventional splitting can produce inefficient Gaussian placement.
MRNF also uses slot recycling to control Gaussian population growth.
When Gaussians become unnecessary and are removed, their allocated positions in the Gaussian pool can be reused for newly generated primitives instead of continuously expanding the underlying tensors.
This has two practical advantages:
As a result, MRNF can perform aggressive local refinement without requiring continuous memory reallocation.
More complex optimization pipeline
Multiple refinement signals and mechanisms must interact during training.
Additional refinement computation
Error and edge analysis introduces processing overhead during refinement stages.
Partially stochastic behavior
Because stochastic noise injection remains part of the strategy, results are not fully deterministic.
High-frequency textures can attract unnecessary refinement
Foliage, image noise, reflections, repeated patterns, or very detailed textures can generate strong edge responses even when additional geometric density is not required.
MRNF is particularly suitable for:
When to choose MRNF
MRNF is a strong general-purpose strategy when you need both robust scene coverage and accurate local detail. It is particularly useful for heterogeneous scenes containing a mixture of large surfaces, fine geometry, edges, and difficult reconstruction regions.
IGS+ is a refinement-oriented strategy designed to improve the precision and efficiency of Gaussian densification.
Unlike MCMC, which allows Gaussian primitives to explore the reconstruction space stochastically, IGS+ follows a more targeted approach: it identifies Gaussians associated with regions that require additional reconstruction capacity and refines those primitives in a geometrically meaningful way.
A central component of the strategy is Long-Axis Split.
Traditional 3DGS densification generally decides whether a Gaussian should be cloned or split based on gradient and scale criteria. Although effective, conventional splitting can introduce child primitives with substantial spatial overlap or place them in positions that do not optimally follow the structure represented by the parent Gaussian.
IGS+ addresses this by taking the anisotropic shape of the Gaussian into account.
Each 3D Gaussian represents an oriented ellipsoid described by its covariance.
The eigenstructure of that covariance determines the Gaussian's principal spatial directions. The axis with the largest spatial extent represents the Gaussian's long axis.
When IGS+ determines that a Gaussian requires refinement, the primitive is split primarily along this dominant direction.
Rather than creating children through unconstrained sampling around the parent, the new Gaussians are positioned according to the parent's orientation and spatial extent.
This has several important effects:
IGS+ also uses image-space information to identify where additional Gaussian resolution is valuable.
Regions containing strong structural edges frequently correspond to areas where a coarse Gaussian representation produces visible reconstruction errors.
By incorporating edge information into the densification criterion, the strategy can prioritize Gaussians associated with:
This allows the Gaussian budget to be directed toward visually significant regions instead of uniformly increasing density across the scene.
Targeted densification is useful only if additional primitives improve the reconstruction.
IGS+ therefore emphasizes efficient Gaussian allocation rather than unrestricted primitive growth.
The objective is to increase local representation capacity where it produces meaningful reconstruction improvements while avoiding redundant Gaussians and excessive overlap.
This generally produces a more compact and spatially organized representation than highly exploratory approaches.
More dependent on the quality of the initial reconstruction than MCMC
Targeted refinement is most effective when the scene is already reasonably represented.
Less exploratory
IGS+ focuses on refining existing reconstruction structures rather than aggressively searching for completely missing regions.
Edge-driven refinement can react to texture rather than geometry
Strong textures, vegetation, noise, or reflections can potentially receive additional Gaussian capacity.
Scenes with very sparse coverage or poor initialization may benefit more from a strategy that includes stochastic exploration.
IGS+ is particularly suitable for:
When to choose IGS+
Choose IGS+ when your input dataset and initial reconstruction already provide good scene coverage and your priority is extracting sharper details, cleaner boundaries, and an efficient Gaussian distribution.
ADC is the densification strategy introduced by the original 3D Gaussian Splatting pipeline.
It manages the Gaussian population using a set of explicit heuristics that determine when existing primitives should be cloned, split, or pruned.
During training, ADC accumulates the magnitude of the view-space positional gradient for visible Gaussians. A high positional gradient indicates that moving the Gaussian would significantly affect the reconstruction loss and is therefore used as a signal that the region may be insufficiently represented.
Gaussians exceeding the densification threshold are processed differently according to their spatial scale.
Small Gaussians with sufficiently high positional gradients are interpreted as representing regions that require additional capacity.
ADC clones these primitives, introducing additional Gaussians that can subsequently optimize independently.
Cloning increases local representation density without initially reducing the spatial extent of the primitive.
Large Gaussians with high positional gradients are interpreted as covering an area that requires a finer representation.
Instead of simply duplicating them, ADC splits the Gaussian into smaller child primitives.
The children can then independently optimize their position, covariance, opacity, and appearance, allowing a large coarse primitive to evolve into a finer local representation.
Not every Gaussian created during training remains useful.
ADC periodically removes primitives that satisfy pruning criteria, particularly Gaussians with very low opacity or primitives that become excessively large.
This prevents the representation from retaining every primitive generated during densification.
The complete process creates a repeating cycle:
Optimize → Measure gradients → Densify → Prune → Continue optimization
Relies heavily on local gradient heuristics
A large positional gradient does not always mean that additional Gaussian density is the optimal solution.
Clone and split operations can introduce redundant primitives
Newly generated Gaussians may overlap significantly.
Density can become uneven
Some regions may accumulate many Gaussians while other difficult areas remain under-represented.
Sensitive to initialization and threshold configuration
Fine details may remain under-reconstructed if the densification signal does not adequately identify them.
Primitive growth can increase model size if densification is aggressive.
ADC is suitable for:
When to choose ADC
ADC is most appropriate for conventional 3DGS workflows with good initialization and predictable scene coverage. For more challenging captures or when higher-quality refinement is required, the newer training strategies generally provide more specialized population-management mechanisms.