[navigation]
Key takeaways
- Banuba Face API runs detection, recognition, and liveness signals on the device, so biometric frames never leave the phone, and there is no cloud round trip mid-onboarding.
- Banuba's engine tracks 68 facial anchor points and reconstructs a 3D face mesh with up to 3,308 vertices via direct 3D reconstruction, not by lifting 2D landmarks into 3D.
- Tracking holds through up to 70% facial occlusion and the full -90° to +90° head-angle range, which stops a check from failing on a user in a car or a badly lit hallway.
- Banuba supplies the signals (head pose, gaze, eye openness, mouth movement, pulse); the fintech team defines the challenge flow, so the risk logic stays in-house.
Why does user verification cost fintech apps so much?
Verification fails in two directions, and both are expensive.
The first is fraud getting through. Presentation attacks use a printed photo, a replayed video, a mask, or a generated deepfake to pass a selfie check, and liveness detection, known in the industry as Presentation Attack Detection, is the layer that catches them. Financial businesses are the most exposed category, which is why crypto exchanges, brokerages, and remote account opening sit at the top of the attack list.
The second is legitimate users dropping out. Every cloud round trip during a KYC check adds latency at the step where the user is already tired of uploading documents, and every face frame sent to a server widens the compliance surface under GDPR and equivalent biometric rules. An Android face detection api that resolves on the handset removes both problems: the check completes on the same screen, and the biometric data has nowhere to travel. On-device versus cloud is the first architectural decision here, and a side-by-side of on-device and cloud face detection SDKs is worth reading first, because the two models diverge on latency, privacy posture, and per-request cost.
How does a face API solve it?
Banuba's engine builds a 3D head model directly from the camera feed. It tracks 68 facial anchor points and reconstructs a face mesh with up to 3,308 vertices, and because the reconstruction is direct rather than a 2D-to-3D conversion, the model does not drift when the input degrades. Tracking stays locked through up to 70% facial occlusion, across the full -90° to +90° head-angle range, and in low light.
That precision turns into usable verification signals: head pose, eye openness, mouth movements, emotion expression, gaze tracking, and pulse detection, plus custom cues such as reflection tests that shine coloured light on the face and read the response in the skin and eyes. Those map onto both liveness approaches.
- Active challenges. Ask the user to blink, smile, or turn their head, then confirm the matching signal fired. Randomizing the sequence is what makes a prerecorded response impractical.
- Passive checks. Read micro-movements, blink cadence, and pulse in the background while the user simply looks at the camera, with no instruction and no added friction.
Banuba does not ship a prescriptive verification flow. It provides the tracking data; the fintech team writes the checks, the thresholds, and the fallback path, so the risk logic stays in-house and auditable while the computer vision comes ready-made. Face recognition sits alongside it: Banuba Face API supports 1:1 matching against a stored reference, the operation behind both document-to-selfie comparison at onboarding and biometric login afterward.
What does the integration look like?
Banuba Face API covers Web, Windows, macOS, Android, iOS, Flutter, React Native, and Unity from one SDK, so an app with a mobile client and a browser fallback integrates once. Runtime requirements are Android 8.0 (API level 26+) or iOS 13+, with a 1280x720 camera recommended.
Teams start with the Banuba Face AR SDK documentation and the public samples: the Android quickstart repo is the shortest path to a running camera feed, and the Android SDK samples repo covers each capability in isolation. A 14-day free trial token gives full SDK access, so the liveness flow can be prototyped on real hardware before any commercial conversation.
Banuba's face tracking in action
Does the engine hold up at production scale?
The clearest evidence sits outside fintech. A SaaS fandom platform built its live video and chat features on Banuba Face AR and passed 1 million monthly active users within two years of launch, serving 224 countries. That deployment is about beauty effects rather than verification, but it stresses the same tracking engine continuously, across the spread of consumer hardware a fintech onboarding flow has to survive.
Ready to test it?
Request a 14-day trial token and run a liveness flow against your own onboarding funnel. Start on the Banuba Face API page.
-
No. A face API handles the biometric layer: confirming a real, live person is present and that their face matches the reference on file. Document validation, sanctions screening, and record-keeping stay separate. Banuba Face AR SDK provides the tracking and liveness signals that layer is built from; the compliance logic around it stays with the fintech team. The liveness detection guide covers how the layers fit together.
-
Face detection locates a face in a frame and answers "is there a face here?". Face recognition identifies whose face it is by matching against a stored reference. Banuba Face API supports both, and most fintech flows use them in sequence: detection to frame and track the user during the liveness challenge, recognition to match the capture against the ID photo. This comparison of face recognition APIs sets out how the main options differ.
-
No. Banuba processes frames on the device, which removes the interception risk of shipping face data to a server and simplifies the GDPR position, since the frames are never transferred. On-device processing is one of the most repeated requirements Banuba hears from regulated buyers, and it is documented on the Banuba Face AR SDK page.
-
Usually, yes. Banuba's tracking tolerates up to 70% facial occlusion and the full -90° to +90° head-angle range, and it is built to hold in low light, which is what keeps legitimate users from being rejected in a car or a badly lit hallway. Banuba's face detection SDK page covers the stress conditions in more detail.
-
Banuba offers two licensing models. A flat annual license covers unlimited usage, so cost stays fixed even if the app scales to a very large user base. An Active User Based model tracks active users over time: significantly lower spend at the start, but a bill that grows with adoption and can overtake the flat fee if the app takes off. The Face AR SDK pricing guide walks through which model suits which growth profile.
