Repository navigation
This update adds configurable face capture, expands platform and language support, and improves inference efficiency, runtime reliability, and SDK distribution.
Highlights
- New Android liveness examples: Add passive liveness and flash liveness algorithm demos to the Android example app.
- Expanded platform SDKs and language bindings: Add Windows x64 CPU/MNN and HarmonyOS ARM64 SDKs to the GitHub Release workflow, including native C/C++ libraries and an ArkTS HAR module. Add Objective-C and Swift APIs with a unified Apple CPU XCFramework package, complete the portable Java/JNI bindings, merge Android JNI into
libInspireFace.so, and support Android 16 KB page sizes. - Best-frame face capture and owned snapshots: Add configurable capture filters for face size, position, stability, pose, quality, sharpness, and brightness, with progress reporting and ranked candidate selection. Reuse owned detection snapshots across processing steps without repeating detection.
- More efficient processing and configurable CPU power: Enable InspireCV Task preprocessing and supported ARM NEON kernels by default; reduce repeated sorting, memory allocation, and tensor copying in detection, tracking, and MNN inference. Add
NORMAL,HIGH, andLOWCPU power modes, withNORMALas the default. - Stronger API reliability and Python usability: Add C API level 2 while retaining the legacy C ABI, strengthen handle and buffer validation, improve FeatureHub search semantics and resource lifetimes, and expand diagnostics and regression coverage. Improve Python context management, exceptions, model downloads, and wheel packaging for Windows, Linux, and macOS.
Supported Platforms and Architectures
The matrix below follows the README's platform coverage. Windows desktop and Windows Server builds currently support x64 with the CPU/MNN backend.
| No. | Platform | Architecture (CPU) | Device / Backend | Supported | Passed Tests | Release Workflow |
|---|---|---|---|---|---|---|
| 1 | Linux (CPU) |
ARMv7 | - | ✓ | ✓ | Workflow |
| 2 | Linux (CPU) |
ARMv8 | - | ✓ | ✓ | Workflow |
| 3 | Linux (CPU) |
x86/x86_64 | - | ✓ | ✓ | Workflow |
| 4 | Linux (Rockchip) |
ARMv7 | RV1109/RV1126 | ✓ | ✓ | Workflow |
| 5 | Linux (Rockchip) |
ARMv7 | RV1103/RV1106 | ✓ | ✓ | Workflow |
| 6 | Linux (Rockchip) |
ARMv8 | RK3566/RK3568 | ✓ | ✓ | Workflow |
| 7 | Linux (Rockchip) |
ARMv8 | RK3588 | ✓ | - | Workflow |
| 8 | Linux (MNN_CUDA) |
x86/x86_64 | NVIDIA-GPU | ✓ | ✓ | - |
| 9 | Linux (CUDA) |
x86/x86_64 | NVIDIA-GPU | ✓ | ✓ | Workflow |
| 10 | macOS | Intel | CPU/Metal/ANE | ✓ | ✓ | Workflow |
| 11 | macOS | Apple Silicon | - | ✓ | ✓ | Workflow |
| 12 | iOS | ARM | CPU/Metal/ANE | ✓ | ✓ | Workflow |
| 13 | Android | ARMv7 | - | ✓ | ✓ | Workflow |
| 14 | Android | ARMv8 | - | ✓ | ✓ | Workflow |
| 15 | Android | x86_64 | - | ✓ | ✓ | Workflow |
| 16 | Android (Rockchip) |
ARMv8 | RK3566/RK3568 | ✓ | ✓ | Workflow |
| 17 | Android (Rockchip) |
ARMv8 | RK3588 | ✓ | ✓ | Workflow |
| 18 | HarmonyOS | ARMv8 | - | ✓ | - | Workflow |
| 19 | Windows | x86_64 (x64) | - | ✓ | ✓ | Workflow |
| 20 | Windows | ARMv8 | - | - | - | - |
| 21 | Linux (Jetson series) |
ARMv8 | Jetson series | - | - | - |
✓ indicates support or testing recorded in the README; - indicates no support or test result claimed in that column. Release workflow links identify SDK packaging pipelines rather than a guarantee that a particular run has completed successfully.
The Apple release combines macOS, iOS, and simulator CPU SDKs into one XCFramework package, with C/C++, Objective-C, and Swift interfaces. CoreML is not included in the release package. HarmonyOS includes native libraries and the ArkTS HAR module, with validation on HarmonyOS hardware still pending. The Android package includes the Java/JNI JAR; a standalone desktop JVM SDK is not a separate Release asset. Python wheels are published separately to PyPI for Windows x64 and Linux/macOS on x86_64 and ARM64.
Features
The following feature matrix follows the README. Backend availability depends on the SDK build and model pack.
| Feature | CPU | RKNPU (RV1109/1126) |
RKNPU (RV1103/1106) |
RKNPU (RK3566/3568/3588) |
ANE (macOS/iOS) |
GPU (TensorRT) |
|---|---|---|---|---|---|---|
| Face Detection | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Landmark | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Face Embeddings | ✓ | ✓ | ✓ | ✓ | - | ✓ |
| Face Comparison | ✓ | - | - | - | - | - |
| Face Recognition | ✓ | ✓ | ✓ | ✓ | - | ✓ |
| Alignment | ✓ | - | - | - | - | - |
| Tracking | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Mask Detection | ✓ | - | ✓ | ✓ | - | - |
| Silent Liveness | ✓ | - | ✓ | ✓ | - | - |
| Passive LivenessPlus | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Flash LivenessPlus | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Face Quality | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Pose Estimation | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Face Attribute | ✓ | - | ✓ | ✓ | - | ✓ |
| Cooperative Liveness | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Face Emotion | ✓ | ✓ | ✓ | ✓ | - | ✓ |
| Embedding Management | ✓ | - | - | - | - | - |
Models and features without NPU or GPU support automatically use the CPU. The ANE column describes optional backend capabilities; the Apple release package contains CPU SDKs only. The Plus labels are retained from the README and indicate edition-specific features.
Resource Package List
Model packs contain the models and configurations for different deployment scenarios and are downloaded separately from the SDK. Update dates below are taken from the README. See the model release for downloads.
| Name | Supported Devices | Note | Last Update | Link |
|---|---|---|---|---|
| Pikachu | CPU | Lightweight edge-side models | Jun 22, 2025 | Download |
| Megatron | CPU, GPU | Mobile and server models | Jun 15, 2025 | Download |
| Megatron_TRT | GPU | CUDA-based server models | Jun 15, 2025 | Download |
| Gundam-RV1109 | RKNPU | Supports RV1109 and RV1126 | Jun 15, 2025 | Download |
| Gundam-RV1106 | RKNPU | Supports RV1103 and RV1106 | Jul 6, 2025 | Download |
| Gundam-RK356X | RKNPU | Supports RK3566 and RK3568 | Jun 15, 2025 | Download |
| Gundam-RK3588 | RKNPU | Supports RK3588 | Jun 15, 2025 | Download |
See the development documentation for integration guides and examples.