Practical overview
When it fits
Use it when you need to study or work with a silent face-liveness model and its documented data-preparation, training, and testing workflow. The demo guidance is for camera-captured faces under specified framing conditions.
Who it is for
- Computer-vision developers exploring or training face anti-spoofing models
- Developers evaluating image-based liveness checks
- Android developers examining the project's separate deployment code and demo
Documented uses
- Train a face anti-spoofing model from categorized image data and prepared face patches
- Test an image and obtain a genuine-or-spoof result with confidence
- Explore an Android demonstration of silent face liveness detection
Requirements & limits
The public repository is open source under the Apache-2.0 license. The README points to an Android APK demo and separate Android deployment source.
- The training pipeline expects three image classes and separate original-image and face-patch preparation.
- The README recommends camera-captured images, a complete face in view, and a face angle below 30 degrees from vertical.
- The project warns that camera model and usage conditions can affect results.
- The high-accuracy model listed in the README is marked as not open source; the APK model is identified as open source.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Prepare data and train a face liveness model
Follow the repository's documented preprocessing and training stages for a three-class dataset.
Before you start
- Python environment with dependencies installed from requirements.txt
- Training images grouped into three class folders
- Original images and face patches prepared at fixed dimensions
- Face detector for producing face patches
Steps
Install the listed Python dependencies and clone the repository.
Arrange the training images into three classes, separating original-image and patch datasets.
Resize original images to fixed dimensions; use a face detector to crop, expand, and resize face patches.
Generate Fourier-spectrum images online during training, then run the documented train.py command with device_ids and patch_info arguments.
Expected outcome
The documented training process uses the classification branch and Fourier-spectrum auxiliary supervision to train a face anti-spoofing model.
Guide limitations
- The instructions do not specify a dataset download or hardware setup.
- The source describes multiple scales and patch preparation but leaves patch_info as a user-provided value.
Test an image with the repository scripts
Run the documented test script on an image using the listed model resources.
Before you start
- Python dependencies installed
- Anti-spoof model files in resources/anti_spoof_models
- Detector model files in resources/detection_model
- A test image in images/sample
Steps
Place the anti-spoof and face-detector model resources in their documented directories.
Place the test image in images/sample.
Run test.py with the image_name argument.
Expected outcome
The test workflow evaluates an image and reports a liveness result and confidence.
Guide limitations
- The project says normal-use test images should be captured by a camera.
- It warns that camera models and usage conditions affect results; a complete face and an angle under 30 degrees from vertical are recommended.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 16 October 2026
- Original website title
- minivision-ai/Silent-Face-Anti-Spoofing · Official homepage title · retained in its source language