Frequently asked questions.

You don't rewrite — you re-point. SightRadar speaks the same request and response shapes as AWS Rekognition, so you keep the AWS SDK you already use and set one endpoint_url. Your existing IndexFaces, SearchFacesByImage, and CompareFaces calls keep working.

Our migration guide shows the exact two-line change side by side with boto3.

A faceprint is a compact numeric summary of a face — the biometric template a recognition system compares, stored by SightRadar as an embedding (a vector of numbers), never as the original photo. Two images of the same person produce faceprints that land close together; different people land far apart. The original image cannot be reconstructed from the faceprint.

More terms are defined in the glossary.

No. SightRadar processes your data only to return your results. We do not pool it with other accounts, do not use it to train general-purpose models, and never sell it. Detect and compare process images in memory and don't persist them; indexed faces live in collections you can delete any time via API or console.

See the trust page and Privacy Policy for the full posture.

Pricing is pay-as-you-go — you only pay for what you use, billed per photo processed (never per face), in USD, with no surprise bills. Real-time index is $0.00093 per photo, batch index is $0.00062 per photo, and search by face ID is $0.00062 per search. Stored faces cost $6 per million per month. There's no subscription and no minimum to start; minimum top-up is $50.

One photo per request, so those rates are what a single API call costs. The exception is /v1/compare: it takes two images and is charged for both ($0.00186 for a two-image comparison).

Run your own numbers on the pricing page calculator.

Compliance with laws like GDPR and Illinois' BIPA is a shared responsibility. SightRadar is built to support lawful, consented use: you are the data controller, we act as your processor, data is isolated per tenant and encrypted in transit, you control retention, and we never train on or sell your data. Our Terms require a lawful basis and, where required, consent — and prohibit unlawful surveillance or covert tracking.

What we can't do for you: obtain consent from your end users or decide your lawful basis — that's the controller's job. This isn't legal advice; see the trust page, Terms, and Privacy Policy.

CreateCollection, IndexFaces, SearchFacesByImage, SearchFaces, CompareFaces, DetectFaces, ListFaces/DeleteFaces, and ListCollections/DeleteCollection. Each maps to a SightRadar endpoint with a deep-linkable anchor in the API reference (e.g. /docs#search-faces).

SightRadar decodes JPEG, PNG, WebP, HEIC/HEIF, TIFF, GIF, and BMP natively — no pre-conversion step. The format is detected from the file's actual bytes, not its extension, so a mislabeled file still works. The one exception is SVG, which is rejected on purpose for security reasons.

Images can be up to 30 MB each. High-resolution photos are automatically downscaled before processing, so full-resolution phone and DSLR shots work without you resizing them first.

This is wider than the competitors: AWS Rekognition accepts only JPEG and PNG (no WebP or HEIC) and caps raw image bytes at 5 MB, Azure Face tops out at 6 MB with no WebP or HEIC support, and Face++ allows only JPG and PNG up to 2 MB — so the WebP images most modern CDNs serve and the HEIC photos iPhones shoot by default work on SightRadar without conversion.

Every search match comes back with a tier that tells you how confident the result is. auto means high-confidence — safe to reveal immediately. review means a plausible match that sits below the auto-reveal bar, so you show it behind a "confirm it's you" step instead of revealing it outright.

Both are real matches above the threshold — the tier only changes how you present them. This matters for a biometric product: it keeps a stranger's photos from being auto-tagged to the wrong person while still surfacing every plausible match for a human to confirm. AWS Rekognition, Azure Face, and Face++ hand you a single raw score and leave that precision-vs-privacy tradeoff entirely to you. A client that ignores the field simply treats every match as auto.

Send an Idempotency-Key header on any billable call and retries are safe: a replay under the same key returns a 409 before any charge, so a network blip or client-side retry can never double-charge you. Without the header, each call is treated as unique — an identical re-run charges again.

Billable calls also return an X-Credits-Remaining header so you can track your wallet balance inline, and when the wallet hits zero, billable calls return 402 Payment Required rather than failing silently.

A photo with zero faces is a valid, successful result — not an error. Index it and you get a clean response with a detected-face count of zero; many APIs throw an error instead, which forces you to write special-case handling.

Before any face is stored, it must clear a quality gate — a minimum bar for size, sharpness, pose, and detection confidence. Faces that fail are rejected rather than stored as a low-quality faceprint, which keeps your collection free of tiny, blurry background faces that would otherwise cause false matches. That's why an index response reports both a detected and a rejected face count. A selfie that fails the gate on search returns a clear reason (for example low_quality_selfie) with no charge for the failed engine call.

Per photo processed, never per face. A photo with five faces costs the same as a photo with one, and a photo with zero faces is still a valid, charged, successful result. Control-plane calls like creating a collection are free.

Every endpoint processes one photo per request, so in practice you are billed per API call. The one exception is /v1/compare, which takes two images (a source and a target) and is charged for both. If you pass one side as a precomputed embedding instead of an image, that side is not a processed photo and is not charged as one.

Consented, lawful identity and recognition workflows — event and wedding photo-tagging, KYC and identity verification, access control, touchless attendance, marketplace and dating trust, and more. Law-enforcement and mass-surveillance use is out of scope and prohibited by our Terms.

See the full list on the use cases page.