Engagement
Training data this specific is not bought off a price list — what it costs depends on the volume, how it is shaped between identities and captures, the conditions you need covered and whether the identities are yours alone. So we scope it against your programme and quote it, and the first step costs nothing: a sample you score yourself.
Four stages, in order. Nobody signs a programme before a sample has been scored, no run starts before its acceptance thresholds are written down, and nothing is remediated before it has been measured. Every package in every category is one of these four — see the packages.
No cost, no commitment
A small set built to your specification — your populations, your capture conditions, the failure you named. You score it on your own detectors and benchmarks, on your own hardware.
Scoped and quoted
The first production-shaped run, against acceptance thresholds agreed in writing before generation starts. If it misses the threshold, that is ours to fix rather than yours to absorb.
Scoped and quoted
Aimed at what the baseline exposed. You bring the confusion matrix or the per-attribute error, and the slice is reweighted to the cells your model actually gets wrong — not spread evenly across a distribution you have already covered.
Annual or multi-year
A standing supply at the cadence your training schedule needs, retargeted each quarter as new frontier models ship and the failure surface moves. Exclusivity and licence terms are settled here.
Packages
Images, then identities × captures each. Select a shape to see what it includes.
| Category | PoC | Baseline | Remediation | Continuous |
|---|---|---|---|---|
| Face recognition | 9,000 1,000 × 9 | 45,000 5,000 × 9 | 180,000 20,000 × 9 | 90,000 / qtr 10,000 × 9 |
| Synthetic detection | 10,000 2,000 × 5 | 100,000 20,000 × 5 | 250,000 50,000 × 5 | 100,000 / qtr 20,000 × 5 |
| VLM training | 10,000 10,000 × 1 | 100,000 100,000 × 1 | 250,000 250,000 × 1 | 100,000 / qtr 100,000 × 1 |
01
No cost, no commitment
IdentityOne person, the same seed every time.
CaptureThat same identity again under different conditions (lighting, angle, age).
ImagesThe total: identities × captures.
On every image, machine-readable beside the file.
Four drivers, and nothing else — no percentage-of-order multiplier, no open-ended hourly line. You will see each of them itemised on the quote with the arithmetic beside it, so the number is reproducible rather than asserted.
How many delivered files the run contains. Larger programmes carry a lower rate per image, so the same shape costs less each at Baseline than at PoC — and a run is never priced worse for sitting slightly above a threshold.
How the volume is split between identities and captures each. The same 100,000 images is a different job as 20,000 identities at five captures than as 100,000 at one — deep runs carry more generation per identity, wide runs carry more curation.
What ships beside each image, which follows the category rather than being bought in grades: match pairs and landmarks for a matcher, attribute fields for a VLM, the condition each image carries for a detector. Extra fields beyond the category's record are quoted as add-ons.
Shared, or exclusive with the seeds reserved to you so the identity pool never appears in anyone else's training set. Exclusivity is the one driver that is about the asset rather than the work.
In writing