Deepfake Detection, Turbo Inference, and Python without GIL — All in 3DiVi Face SDK 3.27
The new version of 3DiVi Face SDK is a standout update that expands core features while delivering a significant performance boost. With a built-in deepfake detection module and support for TensorRT, OpenVINO and no-GIL Python, it’s primed to provide fast, attack-resistant face recognition for your projects from day one.
Key Updates
Deepfake Detection: No Longer Optional for 2025 Biometric Security
Deepfakes aren’t just hype anymore — they’re a real and growing threat to face biometrics worldwide. As the technology becomes easier to access, fraudsters are using it more to spoof identities. To fight back, we’ve added a new Processing Block to our SDK: DEEPFAKE_ESTIMATOR. If you’re involved in KYC, digital onboarding, or cybersecurity, this module is now indispensable.
TensorRT: Tap Into the Full Potential of Your NVIDIA GPU
All Processing Block API modules that can run on NVIDIA GPUs now support inference through TensorRT. This isn’t just optimization — it’s the absolute peak performance you can get from your graphics card. Enable TensorRT and instantly speed up your entire face recognition pipeline — no extra tuning or third-party tools needed.
OpenVINO: Up to 50% Faster Face Recognition on Intel CPUs
Full support for the OpenVINO framework is now integrated into all main Processing Blocks of 3DiVi Face SDK. If you’re using an Intel processor with AVX512/VNNI instructions, your face recognition pipeline can run up to 50% faster — all without switching to a GPU.
We’ve released a new version of the LIVENESS_ESTIMATOR Processing Block (2d_ensemble variant), delivering a 30% improvement in spoof attack detection accuracy compared to the previous version. This makes face anti-spoofing even more robust in real-world conditions — whether you're dealing with printed photos, replay attacks, or masks.
Python API without GIL — Better Parallelism for Developers
Python developers will appreciate this: the Processing Block API now supports -no-gil mode. That means you can efficiently parallelize your face recognition pipelines for better system performance and responsiveness.
Saying Goodbye to CUDA 10.2
We’re officially dropping support for CUDA 10.2, which was released nearly six years ago. The SDK now runs on CUDA 11.x and 12.x — time to move forward with the latest tech.
The new release is available — Upgrade to use faster, more scalable, and flexible face recognition for your projects.