Boxes and masks
2D bounding boxes and per-instance segmentation masks, derived from the engine’s instance ID buffer. Exported as YOLO and COCO.
About
NameFrame is synthetic data for computer vision in Unreal Engine.
It turns controlled Unreal Engine scenes into reproducible computer-vision datasets, with ground truth taken from the engine itself rather than from a model or a human tracing outlines: bounding boxes, instance masks, depth in metres and the camera that rendered every frame.
- Stage
- Commercial Preview
- Engine
- Unreal Engine 5.8
- Platform
- Windows
- Contact
- hello@getnameframe.com
The problem it exists for
Vision models fail on the data nobody has: the rare event, the unsafe situation, the viewpoint no one thought to film, the class that appears twice a year. Collecting those frames in the real world is slow when it is possible at all, and annotating them afterwards is expensive and inconsistent — a person tracing a mask is guessing at object boundaries that a renderer already knows exactly.
A game engine has that knowledge. It knows which pixels belong to which object, how far away each one is, where the camera was and what it was pointed at. The work NameFrame does is turning that into datasets a trainer can read, and then proving the result is correct rather than asserting it.
What it actually produces
One run writes all of this from the same frames, so the modalities agree with each other by construction.
Metric depth
The depth buffer in metres as float32, not an 8-bit greyscale picture of depth.
Camera and metadata
Pose, intrinsics and field of view per frame, plus the spawn manifest that says what was placed where and why.
Its own report
Every run is graded and validated, and the report ships with the data — including the checks it fails.
Where it is right now
NameFrame is in Commercial Preview. It runs as an Unreal Engine 5.8 plugin with a pipeline around it on Windows, and it is verified against that version. It is not a self-serve product yet: teams come in through a private pilot, a few at a time, because a first dataset usually needs a conversation about classes, conditions and volume before anything is generated.
What is public today is the output. Two captures are downloadable with their frames, labels, masks, metadata and validator reports — the airbase capture, graded A · 92.4/100, and the barnyard capture, which scored a C and is published anyway because the reasons it lost points are worth reading. The quality page shows every gate and verification check from a real run, including the one that failed.
There is one thing this site deliberately does not claim: that training on NameFrame data improves accuracy on your real test set. That is a benchmark, it has not been run end to end yet, and until it has, publishing a number would be the one dishonest thing on a site whose whole argument is measurement.
Contact
Questions, pilot enquiries, dataset requests and anything you found wrong in the published data: hello@getnameframe.com.
If you are evaluating this, the fastest useful thing you can do is download a dataset and open it. It costs nothing and it answers more than this page does.
Frame the world. Name the frames.