.png)
3D Gaussian Splatting vs Photogrammetry vs NeRF: Which Method Fits Film, Games and XR?
Updated: 4 days ago
What happens when a production needs the realism of photography, the flexibility of 3D, and the speed of real-time rendering?
That question is driving growing interest in 3D Gaussian Splatting.
Unlike traditional photogrammetry, polygon-based 3D assets, or NeRFs, Gaussian splatting represents captured scenes using spatial Gaussian primitives that can be rendered efficiently in real time.
This makes it promising for film, VFX, games, XR, virtual production, cultural heritage and digital twins.
However, Gaussian splats are not a replacement for conventional CG geometry. They lack many production features such as clean topology, UVs, material controls, collision geometry and skeletal deformation.
The key question is therefore not which technology is “better,” but which scene representation best fits the production task.
This comparison explores that choice across modern film, VFX, games and real-time content pipelines.
Table of Contents
What Is 3D Gaussian Splatting?
3D Gaussian splatting is a scene reconstruction and rendering technique that represents captured environments using large numbers of anisotropic Gaussian primitives rather than relying entirely on polygonal surfaces.
Each Gaussian can be thought of as a small spatial volume with properties describing its position, orientation, size, opacity and colour appearance.
During rendering, those primitives are projected into screen space and blended together to recreate the captured scene.
This is important because the representation is explicit enough to render very efficiently while still preserving many of the view dependent visual qualities associated with neural radiance field methods.
The technique became widely known following the 2023 research paper 3D Gaussian Splatting for Real Time Radiance Field Rendering by Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler and George Drettakis. Their method used 3D Gaussians with an optimized rasterization process to achieve high quality novel view rendering at interactive rates.
In practical visual terms, the result can resemble a navigable three dimensional photograph.
A captured room, street, set or landscape can retain subtle lighting variations, foliage, reflections and fine detail that may require significant cleanup when reconstructed as conventional geometry.
That characteristic is one reason the method has attracted attention from virtual production, VFX and immersive content teams.
It is also important to distinguish visual reconstruction from asset construction.
A Gaussian scene can look remarkably convincing while still lacking the polygon topology required for many traditional modeling, rigging and animation operations.
How Does Gaussian Splatting Work?
A typical Gaussian reconstruction begins with overlapping photographs or video frames captured from multiple viewpoints.
Camera positions must first be estimated so that the reconstruction system understands where each source image was taken.
From there, the optimization process creates and adjusts a large population of spatial Gaussians.
Each primitive develops properties that help reproduce the captured appearance.
These commonly include:
Spatial position
Three dimensional scale
Orientation
Opacity
Colour information
View dependent appearance coefficients
During training, these parameters are repeatedly adjusted until rendered views align closely with the source imagery.
The system can also add, split, remove or refine primitives as it determines where more scene detail is required.
At render time, visible Gaussians are projected toward the camera and composited according to their depth and transparency.
This differs significantly from a NeRF renderer, which traditionally evaluates a neural function repeatedly along camera rays before integrating colour and density values into each pixel. The original NeRF formulation represents a scene through a continuous five dimensional function that maps spatial position and viewing direction to density and emitted radiance.
The explicit primitive structure is a major reason splatting can achieve fast interactive rendering.
Gaussian Splatting vs Photogrammetry
The difference between photogrammetry and Gaussian splatting begins with what each system is trying to reconstruct.
Traditional photogrammetry aims to recover geometric structure.
A conventional workflow may estimate camera poses, calculate a dense point cloud, reconstruct a polygon surface and project photographic textures onto that geometry.
The end product can therefore become an ordinary 3D asset.
Artists can retopologize it, clean the mesh, generate UVs, rebuild materials, create collision, sculpt missing areas and integrate the result into a conventional DCC or game engine workflow.
That geometry centric approach is especially valuable when the captured object must eventually behave like a controllable CG asset.
For productions building characters, props or environments for games, the distinction matters. A captured reference is often only the beginning. It may still need optimized topology, material reconstruction, rigging and level of detail work before entering an interactive pipeline. Mimic Productions addresses that part of the workflow through its 3D modeling for gaming services, where final geometry must satisfy both visual and runtime requirements.
Gaussian reconstruction has a different strength.
Instead of forcing every visual feature to resolve into a clean polygonal surface, it can represent complex appearance through overlapping volumetric primitives.
This can help with scene elements that are awkward for ordinary reconstruction, including fine foliage, small wires, soft edges, reflective regions and certain forms of view dependent appearance. Polycam similarly notes that splatting can handle thin details and reflective subjects effectively, while also warning that poor capture conditions can introduce floating artifacts.
When Photogrammetry Is Stronger
Photogrammetry remains particularly effective when a production needs:
A genuine polygon mesh
Measurable surface geometry
UV based texturing
Collision
Retopology
Sculpting access
Rigging
Physics interaction
Conventional asset export
Predictable integration into established DCC tools
This is why scanning and photogrammetry remain fundamental in character production.
A person can be photographed from many calibrated viewpoints, reconstructed, cleaned and converted into production geometry that can later receive a skeleton, deformation system, facial rig and animation controls.
For XR character work, 3D body scanning for immersive production can provide the geometric foundation required when the result must eventually become an editable, controllable digital human rather than only a captured visual volume.
When Gaussian Splatting Is Stronger
Splat based reconstruction becomes especially attractive when the main priority is preserving captured appearance with minimal geometric abstraction.
Examples include:
Virtual location capture
Set documentation
Digital scouting
Immersive environments
Cultural heritage visualization
Architectural walkthroughs
Difficult foliage
Complex natural environments
Rapid spatial previews
Photoreal background volumes
The crucial distinction is simple.
Photogrammetry is usually more geometry oriented.
Gaussian splatting is usually more appearance oriented.
Gaussian Splatting vs NeRF
NeRF and Gaussian reconstruction share an important objective: both can generate novel camera views from a set of captured images.
Their internal representations are very different.
A Neural Radiance Field models the scene as a continuous function. In the original method, the network receives spatial coordinates and viewing direction and predicts density and view dependent radiance. Rendering requires sampling that function along camera rays.
Gaussian reconstruction instead stores a scene using explicit spatial primitives.
That distinction affects rendering.
Because a splat renderer can project these primitives directly, it generally has a more direct path toward interactive rasterization than the classic NeRF formulation.
This is why the Gaussian splatting versus NeRF discussion often focuses on rendering performance rather than visual quality alone.
Both approaches can reproduce photorealistic novel views.
Their production behavior is not identical.
Which Renders Faster?
In general, 3D Gaussian splatting renders faster than the original NeRF approach.
Classic NeRF rendering requires repeated neural network evaluations along many rays. Gaussian rendering projects explicit primitives and blends them through a specialized rasterization process.
The original 3DGS research was specifically designed to combine high visual quality with real time rendering.
Modern NeRF derivatives have also become substantially faster, so production comparisons should not treat all radiance field systems as one fixed implementation.
The broader distinction remains useful:
Classic NeRF uses an implicit neural scene representation.
Gaussian splatting uses an explicit collection of optimized spatial primitives.
For applications requiring frequent camera movement, interactive exploration or headset rendering, that explicit representation can be significant.
Detailed Comparison: Gaussian Splatting vs Photogrammetry vs NeRF
Factor | Gaussian Splatting | Photogrammetry | NeRF |
Representation | 3D Gaussian primitives | Point cloud, polygon geometry & textures | Neural radiance field |
Primary strength | Photoreal novel views with efficient rendering | Recovering explicit geometry | High-quality neural view synthesis |
Rendering speed | Often suitable for real-time rendering | Depends on optimized mesh complexity | Classic implementations are more computationally demanding |
Editable mesh | Not inherently | Yes | Not inherently |
UV workflow | Not inherent | Standard production workflow | Not inherent |
Capture complexity | Moderate | Moderate to high | Moderate |
Dynamic subjects | Emerging through 4D techniques | Difficult with conventional capture | Possible through dynamic neural-field research |
Rigging suitability | Limited without conversion/hybrid processing | Strong once geometry is cleaned | Limited without additional processing |
Game pipeline fit | Requires specialized renderer or conversion workflow | Strong after optimization | More specialized |
XR potential | Strong for captured spaces & immersive viewing | Strong for interactive geometric assets | Strong for novel-view research & specialized experiences |
Visual realism | Very high for captured appearance | High when geometry & textures reconstruct cleanly | Very high for novel views |
Geometry accuracy | Not its primary design objective | Potentially strong with controlled acquisition | Not its primary design objective |
Production maturity | Rapidly developing | Mature | Established research ecosystem with evolving production tools |
The comparison shows why selecting a technique by image quality alone is misleading.
A beautifully reconstructed environment may be unsuitable for a production if the next department expects editable polygons, UV coordinates, collision volumes or skeletal deformation.
Conversely, constructing a perfectly optimized polygon environment may be unnecessary if the objective is simply to preserve a real location for free camera exploration.
Rendering Speed and Real Time Production
Rendering performance is one of the strongest arguments for Gaussian splatting.
The technique was developed specifically around efficient rasterization of optimized Gaussians, allowing photoreal captured scenes to be viewed interactively.
Recent research continues to explore real time display of Gaussian based radiance fields in demanding display contexts. NVIDIA researchers, for example, published work in 2026 on real time visualization of radiance fields for light field displays, demonstrating the continued interest in efficient spatial rendering pipelines.
Production performance still depends on several variables.
These include:
Number of Gaussians
Screen resolution
GPU architecture
Memory bandwidth
Sorting strategy
Scene complexity
Compression
Level of detail
Target frame rate
Stereo rendering requirements
A desktop visualization and a standalone XR headset do not have the same performance budget.
For this reason, the capture method cannot be separated from deployment.
A scene intended for a high specification workstation may retain a much denser representation than an experience targeting portable hardware.
This is where production planning becomes important.
Mimic's real time asset workflows for XR focus on the same larger problem: visual assets must be prepared around the constraints of the device, renderer and user experience rather than judged only by their source quality.
Can Gaussian Splats Be Used in Unreal Engine?
Yes.
Gaussian splats can be rendered inside Unreal Engine through third party plugins and custom rendering integrations.
As of September 2026, Gaussian splatting is not a native built in Unreal Engine scene representation. An Epic Developer Community response from October 2025 stated that built in support was not currently planned at that time, although third party solutions were already available.
The ecosystem has continued to expand.
Current plugins can import common splat formats such as PLY files and render Gaussian scenes through custom GPU pipelines. Some tools also expose scene cropping, editing, Sequencer controls and 4D playback.
This makes Unreal useful for:
Virtual production tests
Previsualization
Location reconstruction
Interactive set exploration
XR prototypes
Captured environment playback
Experimental volumetric characters
It does not mean a splat behaves exactly like an ordinary Unreal Static Mesh.
Traditional engine systems such as collisions, skeletal animation, mesh deformation, vertex editing and material authoring are built around different asset structures.
Hybrid workflows are therefore likely to remain important.
A production might use a splat for a captured environment while conventional assets provide interactive props, characters, collision geometry and gameplay elements.
Can Gaussian Splatting Capture Moving People?
Static 3D Gaussian reconstruction assumes that the scene remains broadly consistent while it is being captured.
A moving performer violates that assumption.
If a person changes pose while photographs are taken from different viewpoints, different source images no longer describe the same spatial configuration. This can generate duplicated features, unstable geometry, blurred structures or other reconstruction artifacts.
Dynamic Gaussian research addresses this problem by adding time to the representation.
This is commonly described as 4D Gaussian splatting.
Instead of representing only the three spatial dimensions, these systems attempt to model how the Gaussian scene changes over time.
Research presented at CVPR 2024 demonstrated a 4D Gaussian framework designed for efficient dynamic scene representation and real time rendering. The published method reported 82 frames per second at an 800 by 800 resolution on an RTX 3090 in its evaluated setup, although practical results depend on scene complexity, representation density and hardware.
This opens interesting possibilities for volumetric human performance.
The workflow should not be confused with conventional motion capture.
Motion capture records movement data that can drive a rigged character.
Dynamic Gaussian capture records the changing visual appearance of the captured subject.
Those outputs solve different production problems.
Are Gaussian Splats Suitable for Animation?
They are suitable for some forms of temporal playback, but they are not a direct replacement for a rigged animation asset.
A conventional animated character typically contains:
Clean polygon topology
Skeleton
Skin weights
Facial deformation system
Animation controls
Corrective shapes
Hair system
Cloth system
Materials
Level of detail variants
Animators can intentionally alter that character.
They can change timing, modify poses, exaggerate expressions, retarget motion, replace performances and direct individual parts of the body.
A static Gaussian reconstruction does not provide those controls.
A 4D capture can reproduce changing appearance through time, but the result behaves more like volumetric performance content than a traditionally rigged CG actor.
For character driven VFX, controlled deformation therefore remains central. Mimic's character animation services for VFX operate within pipelines where performances can be adjusted, cleaned, retargeted and integrated with the rest of a shot.
Gaussian methods may increasingly complement those workflows, particularly for reference, volumetric playback or captured appearance.
They do not currently remove the need for rigging when direct animation control is required.
Can Gaussian Splatting Replace Traditional 3D Modeling?
Not across a complete production pipeline.
It can replace some reconstruction tasks.
It can reduce the amount of manual scene building required for certain captured environments.
It can preserve visual complexity that would be expensive to reproduce manually.
It can also produce convincing spatial representations quickly enough to change location capture, previsualization and environmental scanning.
But traditional modeling solves problems that appearance capture does not.
A production mesh can be:
Reshaped
Retopologized
UV mapped
Rigged
Simulated
Subdivided
Boolean edited
Used for collision
Exported into established asset pipelines
Modified at component level
Art directed beyond the original capture
Gaussian representations are not inherently designed around those operations.
This is why the strongest near term production model is likely to be hybrid.
A splat can preserve a real location.
Polygon assets can provide interaction.
A scanned performer can supply physical reference.
A clean digital double can provide riggable geometry.
Motion capture can provide editable performance.
Rendering can then combine these different representations according to the needs of the shot or experience.
Applications of Gaussian Splatting in Film, Games and XR
Virtual Production
Virtual production often needs convincing spaces before the final environment has been completely rebuilt.
A spatial capture can provide an immediate visual representation of a location for camera planning, lighting reference, scouting or previsualization.
Rather than treating the scan as the final production asset, teams can use it as an intermediate representation.
This can help departments answer practical questions earlier.
Where can the camera move?
How does a location feel at different focal lengths?
Which structures need full CG replacement?
Which portions may survive as captured backgrounds?
For final pixel VFX work, conventional asset preparation and rendering remain important, particularly when elements must interact with lighting, simulation or compositing. Mimic's VFX rendering services sit at that later stage where assets must hold up under controlled shot production.
Games
Games introduce stricter demands for interactivity.
A playable level needs navigation, collision, dynamic lighting, physics, gameplay triggers and predictable performance.
Gaussian environments can serve as captured scenery or research driven world representations, but conventional meshes remain easier to integrate with many established gameplay systems.
Potential use cases include:
Background environments
Rapid environment capture
Reference for environment artists
Scanned locations
Interactive museum spaces
Experimental narrative experiences
The technology becomes more useful as engine integration, streaming, compression and editing tools improve.
Extended Reality
XR is one of the most natural application areas.
Immersive media benefits directly from realistic spatial representations because the viewer is free to inspect a scene from changing positions.
A captured room, historic site or real environment can therefore become an explorable space without reconstructing every visible feature as manually authored geometry.
The main challenge is performance.
Headsets require stable frame rates, low latency and often stereo rendering.
Dense scenes may therefore need aggressive optimization, streaming or region based loading.
Digital Twins
The phrase digital twin covers several different asset types.
A visually faithful scene representation may be highly valuable for remote inspection, documentation, simulation reference or immersive visualization.
For engineering use, however, visual similarity is not the same as metric or semantic structure.
A Gaussian representation does not automatically know that a visible region is a pipe, wall, switch, door or machine component.
When the application depends on precise measurements, physical properties or semantic object relationships, geometric and structured data may still be required.
Cultural Heritage
Heritage capture is particularly interesting because visually complex spaces often contain surfaces that are expensive to model manually.
A Gaussian reconstruction can preserve the appearance of architecture, artifacts, vegetation and surrounding context for interactive viewing.
Photogrammetry remains important when geometric documentation and measurement are required.
The two methods can therefore complement each other rather than compete.
Previsualization
Previsualization does not always require a perfectly cleaned final asset.
What matters is often spatial understanding.
A captured representation can quickly give directors, cinematographers and VFX supervisors an explorable approximation of a location.
Camera positions can be tested before the complete digital environment is ready.
That can make scene capture valuable long before final rendering begins.
Benefits of Gaussian Splatting
High Visual Fidelity
The technique can preserve complex captured appearance without forcing all visual information into conventional surfaces and textures.
This is useful for natural environments, foliage and other visually dense scenes.
Fast Rendering
Efficient projection and rasterization make interactive rendering one of the defining strengths of the method.
Compared with the original NeRF architecture, splats provide a substantially more direct rendering representation.
Efficient Scene Acquisition
The source material can often come from ordinary overlapping photography or carefully captured video.
This makes the acquisition process accessible, although professional quality still depends heavily on coverage, exposure consistency, sharpness and camera motion.
Strong Novel View Synthesis
Once reconstructed correctly, a viewer can move beyond the exact camera positions used during capture and see convincing intermediate perspectives.
Preservation of Difficult Visual Features
Fine geometry and view dependent appearance can sometimes survive more convincingly than they do in straightforward mesh reconstruction.
Strong Potential for Immersive Content
Fast rendering and photographic appearance make the representation especially interesting for spatial computing, XR and interactive scene viewing.
Production Limitations of Gaussian Splatting
The visual results can make the method appear more universally production ready than it currently is.
Several limitations matter in professional pipelines.
No Native Production Mesh
The representation does not inherently contain a clean polygon surface.
That complicates tasks that depend on topology.
Limited Direct Editing
Editing individual scene structures can be harder than modifying conventional geometry.
Selection and cropping tools are improving, but they do not yet provide the same depth of manipulation as mature polygon modeling packages.
Rigging Is Not Inherent
A splat representation does not naturally contain a skeleton, deformation weights or facial controls.
Character animation therefore requires additional representations or specialized research workflows.
Collision Requires Additional Data
Game engines generally need explicit collision structures.
Visual density does not automatically translate into useful physical geometry.
Memory Can Become Significant
A detailed scene may contain millions of Gaussian primitives.
Rendering may be fast, but storage, VRAM use, streaming and scene transfer still require attention.
Capture Quality Matters
Motion blur, inconsistent exposure, insufficient coverage and shallow focus can produce unstable reconstructions.
Polycam specifically identifies motion blur and shallow depth of field as possible causes of floating artifacts in Gaussian captures.
Unseen Areas Still Need Information
No reconstruction system can reliably recover a hidden surface that was never adequately observed.
A production capture should therefore be planned around coverage rather than simply recording a casual walkthrough.
Dynamic Capture Is More Complex
4D techniques are developing quickly, but dynamic performers require a more sophisticated temporal pipeline than static scene capture.
Engine Integration Is Still Evolving
Support exists in Unreal Engine, Unity and specialist viewers, but splats do not yet behave as universally as conventional polygon assets.
Production teams should validate import formats, rendering features, post processing compatibility and hardware targets early.
Which Method Fits a Production Pipeline?
There is no universal winner.
The correct choice depends on what needs to happen after reconstruction.
Choose Gaussian Splatting When
Photoreal captured appearance is the primary requirement
The scene needs interactive viewing
A clean mesh is not essential
The environment contains difficult fine detail
Rapid location digitization is valuable
The output is intended for spatial viewing or visualization
The pipeline supports a specialist renderer
Choose Photogrammetry When
Surface geometry matters
Measurements are important
The asset needs standard topology based processing
Artists must rebuild or retopologize the scan
Collision is required
The asset will be rigged
The result needs conventional engine integration
Choose NeRF When
Neural novel view synthesis is central to the research or application
The project is built around neural rendering methods
Continuous scene functions offer a specific technical advantage
Production constraints permit the required rendering architecture
Use a Hybrid Pipeline When
Many serious film and interactive productions will benefit from combining representations.
For example, a location may be captured as both a Gaussian scene and a photogrammetric reconstruction.
The splat preserves appearance for visualization.
The mesh provides collision and scene interaction.
Hero objects receive manually cleaned geometry.
Characters remain conventionally rigged.
Performance capture drives animation.
Each representation performs the job it handles best.
That approach is more useful than asking one reconstruction technology to solve every stage of the pipeline.
The Future of Gaussian Splatting
The most important development is not simply higher image quality.
It is the movement toward editable, compressible and temporally aware scene representations.
4D Gaussian splatting is already extending the concept into dynamic content. Research continues to investigate efficient temporal deformation, native four dimensional primitives and compact scene representations for moving subjects.
For production, several areas are likely to matter most.
Better Editing
Artists need stronger tools for selecting, deleting, masking, separating and modifying individual parts of captured scenes.
Improved Compression
XR and streamed experiences need much smaller scene representations without unacceptable visual loss.
More Mature Engine Integration
Deeper Unreal Engine and Unity workflows could make splats behave more naturally beside conventional assets.
Hybrid Geometry Reconstruction
A useful production system may increasingly provide both appearance rich Gaussian content and extracted geometry from the same acquisition.
Dynamic Human Capture
Temporal reconstruction may develop into a practical form of volumetric performance capture.
This will not necessarily replace skeletal motion capture.
Instead, it may create a second category of performance asset: a directly captured dynamic appearance volume.
Semantic Scene Understanding
Future pipelines may combine radiance field style representations with object recognition and scene segmentation.
Instead of treating the environment as a visually coherent field alone, tools could identify meaningful components that artists can select and manipulate individually.
Frequently Asked Questions
What is 3D Gaussian splatting?
3D Gaussian splatting is a scene representation and rendering method that reconstructs captured spaces using many three dimensional Gaussian primitives. Each primitive stores properties such as position, scale, orientation, opacity and appearance. The collection can be rendered efficiently to produce photoreal novel camera views.
How does Gaussian splatting work?
The process usually begins with overlapping photographs or video frames and estimated camera positions. An optimization process creates and refines a collection of Gaussians until rendered images reproduce the source views. During playback, those primitives are projected into screen space and blended to create the final image.
What is the difference between Gaussian splatting and photogrammetry?
Photogrammetry normally reconstructs explicit surface geometry and textures. Gaussian reconstruction focuses on reproducing visual appearance through spatial primitives. Photogrammetry is generally better when a production needs an editable mesh, while splatting is attractive when captured realism and novel view rendering are the priority.
Gaussian splatting versus NeRF, which renders faster?
3D Gaussian splatting generally renders faster than the original NeRF architecture because it uses explicit spatial primitives that can be projected and rasterized efficiently. Classic NeRF systems evaluate a neural field repeatedly along camera rays. Modern neural rendering methods vary considerably, so exact performance depends on the implementation and hardware.
Can Gaussian splats be used in Unreal Engine?
Yes. Third party Unreal Engine plugins can import and render Gaussian scenes, including PLY based captures. Some current tools also provide editing, cropping and dynamic 4D playback. Native built in engine support is not currently standard, so production teams should validate specific plugin requirements before committing to a pipeline.
Can Gaussian splatting capture moving people?
Static splatting is designed primarily for scenes that remain consistent during acquisition. Dynamic variants known as 4D Gaussian methods introduce temporal information and can reconstruct moving subjects. These systems remain an active development area and should be distinguished from ordinary skeletal motion capture.
Are Gaussian splats suitable for animation?
They can support captured temporal content through 4D techniques, but they do not inherently provide the skeleton, skin weights and animation controls of a rigged character. For directed character animation, conventional geometry and rigging remain more controllable.
Can Gaussian splatting replace traditional 3D modeling?
Not generally. It can replace or accelerate certain scene reconstruction tasks, especially when preserving captured visual appearance is the priority. Traditional modeling remains necessary when assets require topology, UVs, deformation, collision, rigging, simulation or extensive art direction.
What are the limitations of Gaussian splatting for production?
Important limitations include the lack of native editable meshes, developing tool support, memory requirements, dependence on good capture coverage, limited conventional rigging support and the need for specialized rendering integrations in many engines.
Conclusion
Gaussian splatting changes an important part of the 3D capture equation.
For decades, production teams have largely treated reconstruction as a route from photographs toward geometry. NeRF demonstrated that convincing three dimensional viewing could instead emerge from an implicit neural scene representation. Gaussian methods now provide another route: explicit spatial primitives capable of reproducing photographed appearance while rendering efficiently.
That combination has clear value.
For virtual scouting, XR environments, digital twins, cultural heritage, spatial documentation and photoreal scene visualization, splats can preserve complex visual information with impressive immediacy.
For game characters, hero VFX assets, animation and physically interactive objects, conventional geometry remains essential.
The distinction is not about which technology wins.
It is about which representation survives the next production step.
When a scene needs to be viewed, a Gaussian representation may be enough.
When it needs to be measured, rebuilt, rigged, simulated, collided with or heavily art directed, a geometric asset remains more practical.
The strongest production pipelines will increasingly combine both.
Comments