3D and Synthetic Media Generation
Generating 3D assets and synthetic imagery for two quite different purposes — visualisation that has to look right, and training data that has to transfer to reality — with honest limits on both.
Two very different jobs share this technology. One is making 3D assets faster for visualisation and product content. The other is generating synthetic imagery to train vision models when real labelled data is scarce. They have almost nothing in common except the tooling, and conflating them produces disappointment in both directions.
3D generation creates or assists in creating three-dimensional assets from text, images or scans. Synthetic media generation produces artificial imagery — frequently rendered from 3D scenes — used to train or augment computer vision models where real labelled data is limited, expensive or impossible to collect.
3D asset creation: useful, with caveats
| Task | Where it stands | Practical note |
|---|---|---|
| Concept and ideation meshes | Strong | Fast exploration; not production topology |
| Photogrammetry and scan clean-up | Strong | Mature and reliable for real objects |
| Texture and material generation | Strong | Genuinely production-useful today |
| Environment and background assets | Good | Where close inspection is not expected |
| Retopology assistance | Improving | Reduces manual work, still needs an artist |
| Production-ready hero assets | Weak | Topology, UVs and rigging still need craft |
| Accurate product geometry | Weak | Use CAD; do not generate what must be exact |
Never generate geometry that must be dimensionally correct
Where a model represents a real product a customer will buy, or anything used for manufacturing, fit or specification, it comes from CAD or a scan. Generated geometry is plausible rather than accurate, and plausible is the wrong standard for anything a purchase decision depends on.
Synthetic training data: the domain gap decides everything
Where it earns its place
Rare events you cannot photograph often enough, hazardous situations you should not stage, defects that occur once a month, scenarios needing perfect labels, and cases where privacy prevents collecting real images of people. In these situations synthetic data is sometimes the only route.
The domain gap is the whole problem
Models trained on rendered images frequently underperform on real ones, because real cameras have noise, motion blur, imperfect focus, unusual lighting and surfaces that do not behave like a renderer's materials. The gap is measured on real held-out data rather than assumed away.
Randomise deliberately
Varying lighting, textures, camera position, backgrounds and distractors across a wide range makes real imagery look like just another variation. This works better than trying to make each render photorealistic, and it is cheaper.
Mix with real data, always
Synthetic data supplements rather than replaces. A modest amount of real labelled data mixed with a large synthetic set generally outperforms either alone, and a purely synthetic model should be treated with suspicion until it is tested on real images.
Validate on real data only
Test sets must be real. A model evaluated on synthetic images is measuring how well it learned the renderer, which is a question nobody needs answered. See object detection and synthetic data generation.
How we work
- Separate the two purposes. Visualisation and training data need different pipelines, different quality bars and different measures of success.
- Establish the accuracy requirement — whether the asset must be dimensionally correct, and therefore whether generation is appropriate at all.
- Build the pipeline into your existing tools, because assets that do not import cleanly into your engine or renderer are not assets.
- For synthetic data, measure the domain gap early on real held-out images, so a poor transfer is a week-three finding rather than a month-four one.
- Keep an artist in the loop for anything customer-facing, because generated assets need cleanup that only a trained eye catches.
- Record provenance and licensing for source scans, reference material and generated assets alike.
How the engagement runs
Domain gap is measured on real data before a synthetic dataset is scaled up.
Purpose and requirements
Visualisation or training data separated; accuracy requirements and target pipeline established.
Pilot
A small asset set or synthetic batch produced; for training data, domain gap measured on real held-out images.
Build
Production pipeline built into your existing tools, with artist review or real-data mixing as appropriate.
Validation
Assets checked in the target engine; models validated on real data only.
Handover
Pipeline, provenance records and the measured limitations handed over.
What you receive
Assets that import cleanly, or synthetic data whose transfer to reality has been measured.
Purpose and requirements assessment
Where generation is appropriate and where CAD, scanning or photography is the right answer.
Asset pipeline
Built into your existing engine or renderer, with artist review for customer-facing work.
Synthetic data pipeline
Deliberate randomisation, real-data mixing and perfect labels.
Domain gap measurement
Model performance on real held-out images, reported honestly.
Validation results
Assets checked in the target pipeline; models validated only on real data.
Provenance and licensing records
For source scans, references and generated assets.
Is this the right engagement?
Worth being direct. 3D and Synthetic Media Generation is the wrong spend in some situations, and those are listed rather than buried.
Good fit if
- 3D asset volume exceeds what your team can model by hand.
- Real labelled images are scarce, hazardous or expensive to collect.
- Rare defects or events need to be represented in training data.
- Privacy prevents collecting real images of people for training.
- Texture, material or environment work is a bottleneck.
Choose something else if
- The geometry must be dimensionally accurate — use CAD or scanning.
- Hero assets need production topology and rigging, which still needs artists.
- Sufficient real labelled data exists and is affordable to obtain.
- The requirement is the vision model itself. See computer vision.
Frequently asked questions
Marked up with FAQPage schema so these answers can surface directly in search results and inside AI assistant responses.
Can AI generate production-ready 3D models?
For concepts, environments, textures and materials, increasingly yes. For hero assets requiring clean topology, sensible UVs and rigging, it reduces manual work rather than replacing the artist. And for anything that must be dimensionally accurate — a real product a customer will buy — use CAD or a scan; generated geometry is plausible, not correct.
Does synthetic training data actually work?
Sometimes very well, particularly for rare events, hazardous scenarios and cases where privacy prevents collecting real images. The determining factor is the domain gap: models trained on rendered images often underperform on real ones. We measure that on real held-out data early rather than discovering it after building a large synthetic set.
How much real data do we still need?
Some, always. Synthetic data supplements rather than replaces — a modest amount of real labelled data mixed with a large synthetic set generally beats either alone, and test sets must be entirely real. A model evaluated on synthetic images is measuring how well it learned the renderer.
How do you close the domain gap?
Mostly through deliberate randomisation — varying lighting, textures, camera position, backgrounds and distractors widely, so real imagery looks like another variation. That works better and costs less than pursuing photorealism in every render. Mixing in real data closes most of what remains.
Can you generate a 3D model of our product?
From CAD or a scan, yes, and that is the right route. We would not generate the geometry of a real product from text or images for anything customer-facing, because the output would be plausible rather than accurate — and a purchase decision should not rest on plausible.
Often paired with this
Most clients combine two or three engagements from the Creative and Media AI pillar. These are the ones that most often run immediately before or after.
AI Image and Brand Asset Generation
A documented visual system, rights cleared before production, and provenance recorded on every asset.
Read more →Digital Humans and AI Avatars
Presenters and avatars on documented likeness consent, always disclosed, with the uncanny valley taken seriously.
Read more →AI Video Generation and Editing
Editing, versioning and localisation at scale, plus short generated sequences with honest limits stated.
Read more →Is this the right engagement?
Tell us what you are trying to build. If a different service fits better, or if you do not need us at all, we will say so.