How to choose a CAD-to-real-time-3D pipeline
A CAD to real-time 3D pipeline turns native engineering data into assets a live 3D application can render at framerate. This is an evaluation framework rather than a product recommendation: what to look at, what to test on your own data, and what to ask a vendor before you commit. It covers what formats a pipeline reads, how it converts exact surfaces into triangles, whether your metadata survives, and what happens when engineering releases a revision. It assumes you have already decided you need a pipeline rather than a viewer or a one-off converter.
Most CAD conversion software comparisons stop at the format list. That is the easy part. The decisions that determine whether your project ships sit further downstream, and almost nobody writes them down.
What a CAD-to-real-time-3D pipeline actually is
A CAD to real-time 3D application pipeline is a repeatable process that ingests native CAD data, converts exact surfaces into triangle meshes, optimizes those meshes against a performance budget, preserves the engineering data attached to them, and delivers deployable assets to a target device. It runs again when the source changes.
Three product categories get confused constantly, and the confusion is expensive.
- A viewer lets someone look at CAD data. It renders on demand and produces nothing you can build on.
- A converter changes a file from one format to another. One file in, one file out, no optimization, no structure.
- A pipeline repeatedly turns source CAD into optimized, organized, deployable assets, with settings you can reuse and govern.
The market is unclear on this distinction. Search for "CAD conversion software" today and the results include a 2D raster-to-vector tool, two generic online file converters, and a roundup that lists AutoCAD, SOLIDWORKS, and CATIA as conversion software. Those are authoring applications. If the top results cannot separate authoring from conversion, buyers may evaluate the wrong category.
If you are still deciding what class of product you need, start with 3D visualization software. This guide begins one step later.
The six stages, and what to evaluate at each
Every credible pipeline does six things. Each stage has its own failure mode.
- Ingest. Read native and neutral CAD, plus reality-capture data. Evaluate format breadth, and whether reading is native or via export.
- Repair and prepare. Heal, stitch, and defeature geometry never built for rendering. Evaluate what the tool fixes automatically.
- Tessellate. Convert exact surfaces into triangles. Evaluate the tolerance parameters you control.
- Optimize and generate LODs. Hit a triangle and draw call budget. Evaluate decimation quality and LOD automation.
- Carry the data. Preserve hierarchy, part numbers, materials, and product manufacturing information (PMI). Evaluate what survives and what is dropped.
- Automate and govern. Run headless, re-run on change, control access. Evaluate the API, the trigger model, and the audit trail.
Ingestion: which formats, and native versus neutral
There are two ways a pipeline can read your CAD, and the difference matters.
Native reading means the pipeline opens CATIA V5, Siemens NX .prt, Creo, SOLIDWORKS, or Inventor files directly, using the vendor's own geometry definition. Neutral reading means someone exports to STEP, IGES, or JT first, and the pipeline reads that.
Neutral formats are useful and widely supported. They are also a problematic handoff. Every neutral hop is a chance to drop assembly structure, part naming, units, custom properties, and product manufacturing information. The hop is also manual, so it does not happen the same way twice and cannot be automated end to end. If your source of truth is a CATIA assembly, reading CATIA natively removes a step, a variable, and a person from the loop.
Reality capture belongs here too. Laser scan and photogrammetry data arrives as point clouds in E57, PTS, or PLY. A pipeline that handles point cloud to mesh conversion alongside CAD lets you put an as-designed model and an as-built scan in the same scene.
Be careful with format counts. "Supports 200 formats" often means 200 extensions, many of them the same format at different versions, some requiring the authoring application installed and licensed on the same machine. Ask which formats read B-rep, which read PMI, and which read metadata. Those three columns are where the count stops being marketing.
Unity Asset Transformer reads more than 70 formats and extensions. Its published table lists native CATIA V4, V5, and V6, NX-Unigraphics .prt from UG11 through NX2412, Creo, SOLIDWORKS, Solid Edge, Inventor, JT, STEP up to AP242, IGES, Parasolid, ACIS, Revit, Navisworks, IFC, and Rhino, plus mesh and point cloud inputs. The same table marks which of those formats carry B-rep, PMI, and metadata, which is the level of detail worth asking any vendor for.
The test to run: send the vendor your ugliest real assembly, not a demo file. Pick the one with 40,000 parts, mixed sources, and a naming convention nobody remembers agreeing to. Check the output for assembly structure, part names, and units.
Outputs matter as much as inputs. Ask which formats the pipeline writes, and map them to where the asset is going. FBX is the common interchange into Unity and Unreal. glTF and GLB are the web and browser formats, and they are what a WebGL or WebGPU application will consume. USD and USDZ are widely used across the mobile AR and DCC ecosystem. OBJ and STL are geometry-only and will drop your hierarchy, so treat them as a last resort rather than a target. A pipeline that writes only one of these limits where you can deploy later.
Tessellation: turning exact surfaces into triangles
Tessellation is the process of approximating an exact CAD surface with a mesh of triangles. CAD describes geometry precisely, as boundary representation surfaces and NURBS. Real-time engines render triangles. Every conversion between the two is an approximation, and you control its quality rather than letting the tool decide it.
Two parameters govern it.
Maximum sag, also called chord tolerance or chord error, is the greatest distance the triangle mesh may deviate from the true surface, expressed in millimeters. Tighten it and you get more triangles hugging the surface. Loosen it and you get fewer triangles and visible faceting.
Maximum angle is the greatest permitted angle between the normals of two adjacent triangles on the same face. Sag alone under-tessellates tight curvature, because a small fillet can sit inside a generous sag tolerance while still looking like a chamfer. Maximum angle forces triangles into those areas.
Published presets show the useful range. Asset Transformer's quality presets run from a maximum sag of 0.01 mm at its densest to 3 mm at its coarsest, with a sag ratio that scales tolerance against the model's bounding box so a bracket and a fuselage are not treated identically. Set sag too coarse and you get faceted cylinders and banding across curved panels. Set it too fine and you generate a triangle count no target device can render.
Tessellation is also where mesh repair happens. Industrial CAD is not built for rendering. Expect to need healing and stitching to close gaps between surfaces, free-edge repair, and normal alignment so face culling does not leave parts of your model invisible from one side.
One sequencing point is worth more than any single setting. Getting the balance right during tessellation beats tessellating densely and decimating afterward. Decimation can only work with the triangles it is given.
Optimization and LODs: working back from your framerate
This is the chain almost nobody draws, and it runs opposite to how most teams work. Start with the device, not the model.
The platform holders publish budgets. The engine vendors largely do not, which is worth knowing before you hunt for an authoritative Unity or Unreal number that does not exist. Unity's graphics optimization documentation is deliberately qualitative, and Khronos moved its glTF asset creation guidelines away from fixed numbers toward per-use-case audit profiles. Your budget comes from your target hardware.
Meta publishes its triangle figures as internally recommended ranges and its draw call figures as example ranges. It also notes that large open-world scenes should target roughly 50% of the stated triangle count, because the published targets came from smaller enclosed levels.
Read the table sideways and the constraint appears. The same assembly that renders acceptably on a desktop review station has to fit inside 250,000 triangles for a shared-space visionOS app, and inside Google's recommended 100,000 triangle limit and 10 MB model size to behave on web and mobile AR. The spread between your loosest and your tightest target is the real design constraint, and for most teams the tightest target is the web, not the headset. That is not a decimation problem you solve at the end. It is a polygon budget you allocate at the start, whether you are building a web configurator, a design review, a training simulation, or a digital twin.
Draw calls deserve separate attention, because industrial data breaks the usual advice. Batching and GPU instancing work well when a scene instantiates a few objects many times. Industrial assemblies have the opposite profile: an enormous number of distinct parts, each with its own material. Merging parts that share a material, and instancing repeated parts such as fasteners, does more for framerate than shaving triangles.
The techniques worth budgeting time for:
- Decimation with sharp-edge preservation. Asset Transformer's decimation documentation shows a CAD part reduced from 140,846 triangles to 13,126, a 90% reduction, with what it describes as almost imperceptible quality loss including on sharp edges.
- Removing occluded interior geometry. An engine will happily render the internal springs of a sealed valve nobody will ever see.
- Defeaturing. Suppress holes, fillets, embossed text, and threads that fall below a pixel at viewing distance.
- Instancing repeated parts. Thousands of identical bolts should be one mesh referenced many times.
- Envelope or proxy generation. Replace a sub-assembly with its outer shell for the furthest LOD tier.
Does your data survive? Metadata, hierarchy, and PMI
Geometry is the part everyone checks. The data attached to it is the part that can quietly disappear. How much that matters depends on the use case: a web configurator may need nothing beyond a part identifier, while a service training application is unusable if hierarchy and part numbers do not resolve on the mesh. Decide what your application actually needs before you score a vendor on this.
A CAD assembly carries more than surfaces:
- Part numbers and identifiers that tie a mesh back to an ERP or PLM record
- Assembly hierarchy, including sub-assemblies, instances, and local transforms
- Materials and colors, assigned at part, body, or face level
- Custom properties, such as supplier, mass, revision, and cost center
- PMI and GD&T, the tolerances, datums, and annotations that carry design intent
Expect three outcomes, not two. Hierarchy and part names usually survive native reading, and often degrade through a neutral export. Materials frequently survive as color but lose their physical definition, so someone reassigns shaders downstream. PMI survives only if both the source format and the reader support it, which is why the format table matters at column level rather than row level.
The test to run: convert a known assembly, then verify that a specific part number still resolves on a specific mesh in the target application. Not "metadata is preserved" in a datasheet. One part, one number, one lookup. Then do it again after decimation, because optimization is a common place for attributes to detach from the geometry they described. For anything touching service, training, or quality, make this a scored requirement rather than a nice-to-have.
Automation, change propagation, and governance
Three criteria that barely appear in vendor comparisons, and the ones most likely to determine whether your pipeline survives its second year.
Automation. Can the pipeline run without a human in the interface? Does it run headless from a command line or an API? Can it process hundreds of files as a batch? Are settings saved as named, versioned profiles, or rediscovered by whoever runs it next? Is there a scripting interface, such as a Python API, for cases the presets do not fit? A pipeline that only runs interactively has exactly one operator, and that operator is a single point of failure.
Change propagation. This is the most expensive thing to get wrong, and it is absent from essentially every competing page. Engineering releases revision C. What happens? In an unmanaged pipeline, nothing happens until someone remembers, so your training application, your configurator, and your review environment all show revision B to people making decisions. Ask three questions. Does a source change trigger a re-run automatically? Does the updated asset reach deployed applications, or only the asset library? How long does a full re-run take, because that is your true update latency. Cloud pipeline automation paired with 3D streaming can close this loop, pushing an updated asset into a running application that prompts users to pull the latest version.
Governance. Converted assets multiply fast, and they are derivative data with a lineage. Look for role-based access control (RBAC), revision history that ties an optimized asset back to its source CAD revision, an audit trail, and a connection to PLM or PDM so the pipeline is not a parallel universe of untracked files. This is where 3D asset management becomes part of the pipeline rather than a separate purchase. This guide to governing 3D data infrastructure covers the architecture layer.
Multi-CAD: normalizing more than one source
Roughly 48% of engineering departments work in a multi-CAD environment, according to a CADENAS PARTsolutions survey. If yours is one of them, normalization is a first-order requirement, not an edge case.
Four things reliably go wrong when several native sources feed one pipeline:
- Units and coordinate systems. One source works in millimeters and Z-up, another in inches and Y-up. Parts arrive at the wrong scale or lying on their side. The error is obvious in one part and invisible in a 30,000-part assembly.
- Naming and material conventions. Each authoring tool exports its own property names. Without a mapping layer, downstream automation that keys on part numbers breaks per source.
- Assembly merging. A CATIA body and an NX bracket have to become one hierarchy with one root, not two scenes side by side.
- Per-source tessellation settings. A tolerance right for a machined housing is wrong for a sheet-metal panel.
Ask whether normalization rules are configurable and reusable, or hard-coded per import.
Build it yourself, or buy a pipeline?
Scripting your own conversion is the most common real alternative, and it deserves an honest hearing.
The case for building. With one source format, one target platform, and stable requirements, a script over an open-source or licensed kernel is a weekend of work and years of low-maintenance service. Teams do this well. You keep control of the output, pay no per-seat licensing, and fit the pipeline to your data rather than adapting your data to a product.
What building costs over time. Format coverage goes first, because sources multiply. Someone acquires a company that runs Creo, and your STEP-only script needs a translator you do not have. Kernel licensing is a real line item, and kernels need upgrading when CAD vendors ship new versions, so every CAD release becomes a maintenance event. Quality is the hardest ceiling: clean sharp-edge preservation and automatic LOD generation is specialist work, and a naive decimator will round your chamfers. The specialist who wrote the script becomes a single point of failure with no documentation and no succession plan.
When building is right. One format, one target, low volume, and geometry expertise already in house. Or a requirement so specific that no product addresses it. NIST quantified the cost of getting interoperability wrong decades ago, finding that imperfect interoperability imposed at least $1 billion a year on the US automotive supply chain, with the majority attributable to time and resources spent correcting and recreating unusable data files. The format count has only multiplied since.
What it costs
Prices go stale, so evaluate models rather than numbers.
- Per seat. A named or floating license for an interactive desktop tool. Predictable, and it scales with operators rather than data.
- Per conversion or per volume. Common in cloud and API pricing. Efficient at low volume, and worth modeling if your source data churns.
- SDK or runtime licensing. For embedding conversion in your own product. Usually negotiated, often with royalty or distribution terms.
- Enterprise or on-premise. Required when data cannot leave your network. Priced as a platform agreement.
Cost is rarely driven by seat count. It is driven by source variety and volume. Ten formats and daily revisions cost more to support than a hundred seats reading STEP once a quarter.
Published pricing is a reasonable proxy for transparency. In the Unity Asset Transformer product range, Asset Transformer Studio lists an annual per-seat price, the Toolkit is bundled into a Unity Industry subscription, and the SDK is quoted. Related tooling such as product configurator software may carry its own licensing, so scope the whole chain rather than the conversion step alone.
A shortlist checklist
Fifteen questions for any vendor, including Unity. An honest one can answer all of them.
- Which of my native CAD formats do you read directly, without an export step?
- For each of those formats, do you read B-rep, PMI, and metadata, or only geometry?
- Which tessellation parameters can I set, and in what units?
- Can I save tessellation and optimization settings as reusable named profiles?
- What is your decimation quality on sharp edges, and can you show a before and after on my geometry?
- Do you generate LOD tiers automatically, and can I set the triangle target per tier?
- Can you remove occluded interior geometry and instance repeated parts automatically?
- What happens to assembly hierarchy, part numbers, materials, and custom properties through the full pipeline?
- Does metadata stay attached to the geometry after decimation and LOD generation?
- Can the pipeline run headless, in batch, from a CLI or API, with no interactive session?
- When source CAD revs, does the pipeline re-run automatically, and does the updated asset reach deployed applications?
- How long does a full re-run of my largest assembly take?
- How are converted assets versioned, and can I trace one back to the source CAD revision?
- What access controls and audit logging do you provide, and do you integrate with my PLM or PDM?
- How does pricing change when I add a source format, double my volume, or need on-premise deployment?
Frequently asked questions
A viewer displays CAD data and produces no output. A converter changes one file into another format, one at a time, with no optimization. A pipeline repeatedly ingests source CAD, tessellates it, optimizes it against a performance budget, preserves its metadata, and delivers deployable assets. Only the third is a production capability.
No intermediate FBX or OBJ step is required. Pipeline tooling reads native formats such as CATIA, NX, Creo, and SOLIDWORKS directly and converts the exact surfaces into meshes during import. FBX and glTF are useful for preprocessing very large assemblies outside the editor, but they are one workflow option, not a prerequisite.
Tessellation tolerance, usually expressed as maximum sag or chord tolerance, is the greatest distance the triangle mesh may deviate from the true CAD surface. Published presets span roughly 0.01 mm for dense output to 3 mm for coarse output. Choose it against your triangle budget and viewing distance, and pair it with a maximum angle setting so tight fillets tessellate correctly.
Neither Unity nor the platform holders publish a recommended LOD count. In practice, three tiers plus a culled state is a common starting point for VR and desktop, and two or three is typical for web delivery where download size binds first. Work backwards from your device budget: if your furthest tier exceeds it, add a tier.
It depends on the format and the reader. Some survives, some degrades, some is lost. Assembly hierarchy and part names usually survive native reading. Materials often survive as color but lose their physical definition. PMI survives only where both the source format and the reader support it. Test with a specific part number rather than trusting a datasheet.
That depends on whether your pipeline is automated. In a manual pipeline, nothing happens until someone notices, and deployed applications keep showing the previous revision. In an automated pipeline, the source change triggers a re-run, and the updated asset reaches the asset library and deployed applications through streaming or an update prompt.
Not for interactive work. Desktop tools and editor plugins are built for technical artists and engineers, and rule-based presets cover most repeatable preparation. You do need developer capability for the automation layer: scripting an API, wiring triggers to your PLM, and running conversions in CI or the cloud.
