Gaussian Splatting BIM: overview, benefits, and IFC relevance

Gaussian Splatting, NuRec, and neural reconstruction are technologies based on artificial intelligence and neural representations of the scene, useful for reconstructing real environments in 3D from images, sensors, and data collected from the field. Their main value is generating a photorealistic, navigable, and visually faithful representation of reality.

In the context of digital twins, these technologies do not replace BIM or the IFC model and belong to a different layer: they reconstruct the visual appearance of the space, while IFC and openBIM describe the technical meaning of the elements. A scene reconstructed with Gaussian Splatting can show walls, windows, systems, or infrastructures, but it does not automatically contain BIM objects, IFC properties, relationships, classifications, or lifecycle data.

For this reason, the most credible future is not “NuRec versus IFC,” but the integration between neural twin, semantic twin, and open workflows where OpenUSD can also play a role in connecting, composing, and simulating 3D scenes. In this perspective, AI does not eliminate BIM: it complements it, enhancing the visual layer of the digital twin and leaving to openBIM the task of providing meaning, structure, and interoperability to the project’s data.

What is Gaussian Splatting in BIM?

Gaussian Splatting is a technique for photorealistic 3D representation that can support BIM workflows, reality capture, and digital twin, but it is not a BIM informational model.

In 3D Gaussian Splatting, a scene is not primarily represented as a mesh or a traditional point cloud, but as a set of gaussian primitives oriented in space, each with position, shape, color, opacity, and other parameters optimized to generate photorealistic views.

In the AEC context, Gaussian Splatting mainly concerns:

The Gaussian Splatting BIM is therefore a useful expression to describe the meeting between neural reconstruction and informational processes, but it does not indicate a new BIM format. A scene generated with Gaussian Splatting can show walls, windows, systems, roads, or industrial environments, but it does not automatically contain IFC classes, technical properties, spatial relationships, or maintenance data.

A gaussian can contribute to the visual representation of a wall, but it is not a IfcWall. It can reconstruct the appearance of a window, but it is not an IfcWindow. It can reproduce a system, but it does not know its function, capacity, classification, or maintenance obligations.

The central point is this: a photorealistic model is not automatically an informational model.

What is neural reconstruction and why does it also relate to BIM?

Neural reconstruction is the set of techniques that allows the reconstruction of three-dimensional scenes from real data, using artificial intelligence, sensors, cameras, LiDAR, and neural representations of space. In the AEC sector, it is relevant because it can transform surveys and real-world acquisitions into photorealistic 3D environments, useful for digital twin, simulations, inspections, comparison with the project, and scan-to-BIM workflows.

Gaussian Splatting fits into this area: it is a 3D representation and rendering technique that allows for photorealistic visualization of scenes reconstructed from reality. In other words, neural reconstruction describes the general process of scene reconstruction; Gaussian Splatting is one of the technologies that can make this scene navigable, realistic, and usable in near real-time.

In this context, NVIDIA Omniverse NuRec is often cited as it represents a concrete case of application of neural reconstruction and 3D Gaussian Splatting to navigable and usable 3D scenes in simulation environments. NuRec is not a BIM tool and does not automatically generate IFC models, but it clearly illustrates the technological direction: acquiring data from the real world, reconstructing a photorealistic 3D scene, and making it usable in OpenUSD-based workflows, simulation, and physical AI.

Why is neural reconstruction relevant for the AEC sector?

Neural reconstruction relates to BIM and digital twin because it enables:

• photorealistic reconstruction of real environments;
• navigation of complex spaces captured by sensors;
• support for simulations and training of artificial intelligence systems;
• comparison of design models and the actual state;
• complementing workflows of surveying, inspection, and scan-to-BIM.

NVIDIA also connects NuRec to the creation of 3D environments for training and testing of physical AI systems, robotics, and autonomous driving; the NuRec libraries also introduce rendering workflows based on 3D Gaussian Splatting for capturing, reconstructing, and simulating the real world.

This makes NuRec interesting for the BIM world, but it does not transform it into a BIM authoring tool.

Neural reconstruction measures and reconstructs what appears; BIM structures what an element is, what properties it has, and how it participates in the processes of design, construction, and management.

Gaussian Splatting BIM

IFC and Gaussian Splatting: why they do not describe the same thing

To understand why Gaussian Splatting, NuRec, and neural reconstruction cannot replace BIM, it is necessary to clarify what an IFC model truly contains. The comparison is not just about two different ways of representing a 3D scene but about two different levels of information: on one hand, the visual appearance of reality, and on the other, the technical meaning of the elements that make up a work.

IFC is the open standard for BIM information exchange: it describes not only 3D geometries but objects, properties, quantities, relationships, and lifecycle data.

The most recent official version indicated by buildingSMART is IFC 4.3.2.0, while IFC 5 is being refactored to enable more advanced use cases.

In an IFC model, a building or infrastructure is not represented as a simple 3D scene. They are described as an information object system:

• a wall can be an IfcWall;
• a beam can be an IfcBeam;
• a space can be an IfcSpace;
• an element can have properties, materials, quantities, and classifications;
• each object can be linked to other objects, systems, levels, or documents.

What can IFC do?

An IFC model can answer questions such as:

• Which walls have a certain fire resistance?
• Which structural elements belong to a certain floor?
• Which components require maintenance?
• What quantities are associated with a category of works?
• Which spaces are served by a system?
• Which objects belong to a specific classification?

This is why the IFC model is a semantic database of the work, not just a simple 3D representation.

NuRec and IFC in comparison: neural twin vs semantic twin

NuRec and IFC do not compete on the same level: NuRec describes how a scene looks, IFC describes what the elements that make up a work mean.

The difference can be summarized like this: IFC encodes the intent and meaning of the work; NuRec encodes the visual measure of reality.

The IFC model answers questions like: “What is this element?”, “What properties does it have?”, “To which system does it belong?”, “How does it connect to other objects?”. NuRec answers a different question: “What is this space made of and how does it look, viewed from here or from another point?”.

These are different questions. Both useful. But not interchangeable.

Gaussian Splatting BIM

Why is OpenUSD the point of contact between Gaussian Splatting and BIM?

OpenUSD, which stands for Open Universal Scene Description, is an open ecosystem for describing, composing, and exchanging complex 3D scenes. It originated in the world of graphics, visualization, and simulation, and allows for organizing geometries, materials, lights, animations, variants, and layers of a digital scene.

This is why OpenUSD is a possible convergence ground between Gaussian Splatting and BIM: it can facilitate workflows between neural reconstruction, simulation, visualization, digital twin, and informative models. More precisely, OpenUSD can become a point of contact in the workflows of composition and simulation of 3D scenes, not in the complete management of BIM semantics.

The collaboration between buildingSMART International and Alliance for OpenUSD, announced on October 1, 2024, aims to explore synergies between open digital standards, IFC, and OpenUSD.

Here the technical convergence arises:

• NuRec produces scenes in USDZ format;
• OpenUSD provides a substrate for representing and composing 3D scenes;
• IFC 5 is evolving to address more advanced use cases;
• buildingSMART and AOUSD have formalized a cooperation path on open digital standards.

But this convergence does not mean that IFC and USD become the same thing. USD is a container and framework for complex digital scenes. IFC is an informational model for built assets. OpenUSD can bring workflows closer, but does not replace BIM semantics.

Where do NuRec and IFC really intersect?

Although belonging to different domains, NuRec and IFC share several operational touchpoints within modern AEC workflows. Both contribute to the construction of the digital twin, but with complementary roles: NuRec helps to capture and accurately represent the existing reality, while IFC organizes and structures the necessary information to design, verify, build, and manage the work over time. Here are the areas where this integration is most evident.

Reality capture and state of the art

NuRec and BIM meet in workflows based on real surveys but produce different results.

NuRec starts from data acquired through sensors, cameras, LIDAR, and multi-sensor recordings. Many BIM workflows also start from reality, especially when working on existing buildings, infrastructure, historical heritage, or complex assets.

Scan-to-BIM arises precisely from the need to transform surveys, point clouds, and acquisitions into informational models. The difference is that NuRec produces a navigable photorealistic scene, while scan-to-BIM aims to construct or update informational objects.

A possible integrated workflow can be read as follows: real acquisition → photorealistic reconstruction → comparison with BIM model → recognition or segmentation of elements → semantic enrichment → updating IFC, CDE, or digital twin.

Digital twin

NuRec and IFC can coexist in a mature digital twin: the former as a visual layer, the latter as a semantic layer.

The neural twin is powerful when it is necessary to see, simulate, and reproduce an environment realistically. It is useful for immersive visualization, training artificial intelligence systems, robotics, autonomous driving, and simulations.

The semantic twin is essential when it is necessary to manage information, properties, classifications, maintenance, controls, computations, authorization processes, and lifecycle data.

A truly useful digital twin should not choose between the two. It should integrate them.

Georeferencing

Georeferencing is the bridge between BIM, GIS, surveying, digital twin, and neural reconstructions.

In territorial, infrastructural, and urban digital twins, visual data and informational data must share a coherent spatial reference. BIM must connect to the geographic context; neural reconstructions must align with reference systems useful for design, management, and analysis.

In this scenario, IFC, GIS, point clouds, and photorealistic scenes can coexist if properly coordinated.

Point clouds and external data

The informational model does not need to encompass everything: it can reference and coordinate external data, such as point clouds, surveys, or reconstructed scenes.
This is a crucial point. In an evolved digital twin, a point cloud, a Gaussian Splatting scene, an IFC model, and a GIS layer can coexist, each with its role.

The value lies not in transforming everything into IFC or everything into USDZ, but in creating reliable links between different data.

Asset harvester: 3D objects derived from reality, not yet BIM objects

Another point of contact between NuRec and BIM concerns the ability to recognize, correct or replace certain elements within a reconstructed scene.

NVIDIA Omniverse NuRec describes workflows that allow you to work on the assets present in USDZ reconstructions, meaning elements of the scene that can be identified, corrected, replaced, or modified. A 3D scene reconstructed from reality is not treated merely as a navigable image, but can begin to contain recognizable parts to intervene on.

This is an important step, as it brings neural reconstruction closer to the world of objects. However, visually recognizing an element does not automatically mean transforming it into a BIM object.

A door recognized in a 3D scene, for example, is not yet an IfcDoor. To become one, it should have technical properties, reliable dimensions, materials, classifications, relationships with walls and spaces, any fire safety data, maintenance information, and connections to the rest of the information model.

Therefore, the asset recognition and management tools described by NVIDIA are interesting, but do not eliminate the role of BIM. They help to make the scene more organized and modifiable, but the leap towards BIM semantics still requires structured, controlled, and interoperable information.

Where NuRec and IFC do not touch: the semantic gap

The main limitation of neural reconstruction compared to BIM is the semantic gap: a photorealistic scene does not automatically know what the objects it shows represent.

A reconstruction with Gaussian Splatting can show:

• a wall;
• a ceiling;
• a road;
• an installation;
• an industrial environment;
• a door;
• a pillar.

But it does not natively distinguish:

• a load-bearing wall from a partition;
• a fire-rated door from a regular door;
• a reinforced concrete pillar from a cladding;
• a fire pipe from a conduit;
• a structural element from a piece of furniture;
• a construction discrepancy from a design choice.

This information does not belong solely to the visual surface. It belongs to the information model.
Saying that “Gaussian Splatting will replace BIM” is therefore misleading. Gaussian Splatting can enhance the way we visualize and reconstruct reality, but it does not replace the data structure necessary to design, compute, verify, approve, deliver, manage, and maintain a work.

Design intent and measurement of reality: what is the difference?

The IFC model represents design and informational intent; the neural reconstruction represents a measurement of the real state acquired at a certain moment.

This distinction is fundamental:

This difference is not a problem. It is an opportunity. The comparison between the IFC model and real reconstruction can enable high-value use cases:

• verification between as-designed and as-built;
• monitoring of work progress;
• interference control concerning the real state;
• survey of discrepancies;
• updating of the digital twin;
• support for maintenance;
• photographic and three-dimensional documentation of the state of places.

In this scenario, NuRec does not replace IFC. It complements it.

Why is editability a decisive difference?

A BIM object can be modified as an informational object; a neural reconstruction must be corrected, replaced, re-optimized, or regenerated.

In a BIM model, it is possible to:

• change the thickness of a wall;
• update the material;
• replace a window;
• modify a property;
• associate a classification;
• link a document;
• update a quantity;
• modify a relationship between elements.

In a neural reconstruction, on the other hand, you do not modify a wall because the wall does not exist as a semantic object. There exists a distribution of visual primitives that contribute to its representation.

To intervene in the scene, different procedures are needed: correction, refinement, asset replacement, re-optimization, or new acquisition. This confirms that editing a BIM model and correcting a neural scene are conceptually different operations.

Where does Gaussian Splatting fit in a BIM digital twin?

To understand why Gaussian Splatting, NuRec, and neural reconstructions do not replace BIM but can integrate it, it is useful to read the digital twin as a structure composed of multiple layers.

From this perspective, a photorealistic reconstruction does not need to contain all the information of the work. It can instead represent the visual layer of the digital twin, that is, what allows you to see the real state of a building, infrastructure, or environment acquired through sensors.

Appearance layer:

Includes:

• surveys;
• images;
• point clouds;
• meshes;
• Gaussian Splatting;
• USDZ scenes;
• photorealistic views;
• visual simulations;
• neural reconstructions.

It serves to see the real world, navigate it, understand it visually, and reproduce it in simulation.

Meaning layer:

Includes:

• IFC;
• openBIM;
• properties;
• classifications;
• quantities;
• relationships;
• systems;
• maintenance data;
• documents;
• constraints;
• responsibilities;
• lifecycle information.

It serves to query, manage, and govern the asset.

The value arises when these two layers interact. A photorealistic reconstruction can show the real state of a building; the IFC model can state that the element is a fire-rated wall, belongs to a compartment, has a certain performance, and must be verified according to a procedure.

Without appearance, the digital twin risks being abstract. Without meaning, it risks being just a visual representation.

OpenUSD and IFC 5: what is already concrete and what is still evolving?

The convergence between OpenUSD, IFC 5, and digital twin is a relevant technical direction but should not be presented as a fully resolved integration.

buildingSMART indicates that IFC 5 is an ongoing refactoring aimed at bringing IFC to a next technical level and enabling new advanced use cases.

The direction of convergence is clear, but the application maturity is still evolving.

For the AEC sector, this is an important moment because it relates three worlds that have often traveled separately until now:

• semantic BIM;
• advanced 3D visualization and simulation;
• neural reconstruction of the real.

The point is not to abandon openBIM. It is to bring it into broader workflows capable of integrating reality capture, AI, GIS, simulation, and digital twin.

What role for ACCA software and for openBIM?

The role of openBIM is to give meaning to data, while neural technologies are increasingly making it simpler to generate the visual layer of the digital twin.

In this scenario, the value does not lie in opposing IFC to neural reconstruction. The value lies in clarifying the picture.

ACCA software, the number 1 expert in IFC openBIM^® and the company with the largest number of IFC certified software from buildingSMART International in the world, naturally fits into the layer of meaning: the one where data is not only visualized but structured, queried, linked, and governed.

On one side, interest in IFC 5, OpenUSD, and USD-based workflows is growing; on the other, the integration between the information model, geospatial data, and surveying remains central, even in environments oriented toward territorial and infrastructural digital twins.

As the production of the visual layer becomes simpler, more accessible, and automated, the competitive advantage will increasingly shift towards the ability to give meaning to data.

A photorealistic scene can be generated by sensors, GPUs, and AI. But deciding what that scene means, how it connects to the project, what information it contains, what responsibilities it activates, and how it fits into the management processes of the work remains an openBIM topic.

To explore how open standards, IFC, and data sharing environments can support interoperable workflows not tied to proprietary formats, discover ACCA software’s openBIM platform.

FAQ – Gaussian Splatting and BIM

What is the difference between Gaussian Splatting and BIM?

Gaussian Splatting is a photorealistic visual representation technique based on neural primitives. BIM is a method of information management of the work based on objects, properties, relationships, and lifecycle data. The former describes how a space looks; the latter describes what it means.

Does NuRec produce a BIM model?

No. NuRec produces a 3D scene reconstructed from reality, with USDZ output, useful for visualization, simulation, and training of physical AI systems. It does not natively produce a semantic IFC model with classes, properties, and BIM relationships.

Can a USDZ file contain BIM information?

USDZ can contain a 3D scene, but this does not automatically mean it contains BIM semantics. The point of contact between USD and BIM is evolving, especially through OpenUSD, IFC 5, and the cooperation initiatives between buildingSMART and AOUSD.

Will IFC 5 use OpenUSD?

It is correct to speak of a direction of convergence, not of fully established integration. IFC 5 is evolving, and buildingSMART has initiated cooperation with AOUSD to explore synergies between IFC and OpenUSD.

Can Gaussian Splatting be useful in scan-to-BIM?

Yes. It can be useful as a visual and reality capture layer, especially to represent the existing state. However, to obtain a queryable BIM model, a semantic step is needed: identifying objects, classifying them, assigning properties, and linking them to an information model.

Why does the IFC model remain important in digital twins?

Because a digital twin must not only show an asset: it must allow for querying, managing, updating, and linking it to the processes of design, construction, maintenance, and management. These functions require structured semantic data.

What is the role of OpenUSD in AEC digital twins?

OpenUSD can become a composition and exchange environment for complex 3D scenes, simulations, and digital twin workflows. In the AEC sector, it can facilitate integration between advanced visualization, simulation, and informational data, but it does not replace the IFC model.

What is the difference between a neural twin and a semantic twin?

The neural twin reconstructs the appearance of reality through sensory data and AI models. The semantic twin describes the meaning of elements through objects, properties, classifications, and relationships. A mature digital twin should integrate both.

You might also be interested in the article “AI construction“