“Synthetic data” covers two industries that share a name and little else. One learns from your databases and produces new rows. The other builds virtual worlds and photographs them. Buyers who start by searching for “synthetic data” often end up on the wrong side of the market for a week before they notice. Here is how to tell which side you need.
Two different starting points
Tabular generation starts from your data. You hand the generator a real table, or a set of linked tables, and it learns the distributions and relationships. Everything it produces is a recombination of patterns it found there. The reason to buy is usually access: the real data is too sensitive to share with a test team, a partner or an outside analyst.
Simulation starts from a model of the world. An artist or a procedural system builds 3D scenes: roads, warehouses, ships, sensors. A renderer produces images or sensor readings with every label known exactly. No real data is needed to begin, though you need it to check the result. The reason to buy is usually coverage and cost: rare scenes, dangerous scenes, and pixel-perfect labels without paying annotators.
How the two compare
| Tabular synthetic data | Simulated images and sensor data | |
|---|---|---|
| Input | Your real tables | 3D scenes, assets and sensor models |
| Main reason to buy | Share or test data that cannot be used directly | Cover rare scenes and get exact labels |
| Main risk | Exposing real records; losing rare values | Domain gap: models trained on renders do worse on real images |
| Key test | Fidelity, utility on real data, privacy | Accuracy on a real validation set, per class and per scenario |
| Vendors on our list | Tonic.ai, Syntho (with MOSTLY AI), Aindo, K2view, SAS Data Maker | Rendered.ai, Parallel Domain |
| Published prices | None of the five, on the pages we read | Rendered.ai from $5,000 a month (Teams) and $15,000 a month (Organizations) |
Vendors in this table may pay us an introduction fee if a team we put in touch signs with them; it does not change what we say about them.
Vendor details are from each vendor’s own pages, read 7 November 2026. The full list, with ownership changes, is on the synthetic data companies home page.
Signs you need tabular generation
- The bottleneck is permission, not volume: legal or security will not let the real table leave a team.
- You need a realistic staging database and masking leaves too much real data behind.
- You want to share data with a partner or a vendor and need a privacy claim you can defend.
- Your data lives in rows and columns: customers, orders, events, inventory.
Tabular vendors differ most in relational support (keeping keys consistent across many tables), where the software runs (inside your cloud account or theirs) and what privacy method they offer. The tabular market has also consolidated: SAS bought Hazy’s software assets in November 2024 for SAS Data Maker, and Syntho bought the MOSTLY AI trademark and related assets in June 2026.
Signs you need simulation
- The bottleneck is examples: the scenes you need are rare, dangerous or expensive to capture.
- Labeling cost dominates: segmentation masks, depth or 3D boxes by hand.
- You work with sensors other than ordinary cameras. Rendered.ai lists SAR, infrared, multispectral and X-ray imagery; Parallel Domain focuses on perception for autonomous vehicles and robotics.
- You want to test a model on scenarios you design, not only on what happened to be recorded.
Simulation vendors differ in sensor realism, how much you can control scene variation, whether they sell a platform you run or finished datasets, and how they help you close the domain gap.
Where the line blurs
A few cases sit between the two. Free text and documents are generated by language models rather than by either tabular generators or renderers; Tonic.ai’s Textual product and its Fabricate tool (bought in April 2025) cover text and schema-driven generation. Time series from sensors can be handled by some tabular tools or produced by simulation. Video of people raises the privacy questions of the tabular world with the rendering methods of the simulation world.
If your project spans both, treat it as two purchases. One vendor that claims to do everything well usually does one of them well.
How to brief each side
For a tabular vendor, describe the tables in words: how many, how many rows, the column types, the links between them, where the software must run and the privacy promise you need. For a simulation vendor, describe the task, the sensor (make and model if you can), the scenes and classes that matter, the label types and the real validation set you will judge results on.
In both cases, decide the test before the trial. The detail is on the synthetic images and simulation page for the rendering side.
Getting the side of the market right first saves more time than any comparison of vendors within it.