insightface

Releases

1. ▸▾2.13 Oct 2026

   Add person detection, face recognition and body ReID through PersonAnalysis, with separate get(), match() and optional update() calls.

   PersonAnalysis

   • Add person detection, face recognition and body ReID through PersonAnalysis, with separate get(), match() and optional update() calls.
   • Support cheetah_s and cheetah_l on CPU and NVIDIA CUDA, with first-use model downloads and configurable face and body detection input sizes.
   • Add a desktop workflow for local video, webcams and RTSP, with reference-photo registration, analysis-rate controls and advanced settings.

   See the PersonAnalysis guide for usage, configuration and examples, or the GUI guide for desktop use.

   Open source →
2. ▸▾2.08 Sept 2026

   Add opt-in RGB liveness through FaceAnalysis(addons=["liveness"]), with normal / observe modes and a configurable live-score threshold (default 0.8).

   Liveness update

   • Add opt-in RGB liveness through FaceAnalysis(addons=["liveness"]), with normal / observe modes and a configurable live-score threshold (default 0.8).
   • Return per-face scores and input-rejection guidance. Normal mode retains detections while skipping recognition for faces that do not pass.
   • Download and verify the separate addon model under <root>/addons/liveness.onnx. See usage and result fields.

   PrivateFrame update

   • Add local video face detection/tracking with Gaussian blur or mosaic through the Python API, insightface-privateframe CLI, and Evaluation Studio GUI.
   • Add all, blur_only, and exempt policies, with a folder of reference photos for selecting people.
   • Export editable analysis JSON and render it again without another inference pass; provide analyze, process, and render workflows.
   • Add configurable analysis sampling (Fast mode, target 15 FPS by default), interpolation between scans, encoding/audio controls, and structured CLI status and progress.
   • Add the PrivateFrame feature and usage guide.

   Runtime and model updates

   • Add manifest-backed raccoon_s / raccoon_l model packages and make raccoon_s the default for new GUI configurations.
   • Add automatic CoreML/CUDA/CPU provider selection and persistent CoreML compilation caches; use reusable fixed-shape SCRFD sessions per resolution.
   • Set ORT_DISABLE_TELEMETRY=1 before importing ONNX Runtime, where supported.
   • Add the privateframe installation extra and include it in the gui extra.