VisionLite

VisionLite is an Android face vision toolkit focused on low latency, low memory use, concurrency, and developer experience.

It ships five face modules:

• Face detection
• Face alignment
• Face embedding
• Face anti-spoofing
• Face quality

All models are bundled with the library so the app only uses the public API.

Each module has its own config and model preset, so you can tune latency, quality, and memory independently.

---

Quick path

1. Add the library dependency.
2. Create VisionLite with the face modules you want to configure.
3. Run detection on ImageProxy, ByteArray, Bitmap, or captured requests.
4. Use face.crop(...), faceAligner, faceEmbedder, and match(...) for enrollment and recognition.

---

Installation

repositories {
    maven("https://jitpack.io")
}

dependencies {
    implementation("com.github.kkokotero:VisionLite:1.0.1")
}

---

Basic usage

val faceDetector = VisionLite.faceDetector(context) {
    mode(FaceDetectionMode.LOW)
    trackingEnabled(true)
    maxConcurrentRequests(2)
}

val faceAligner = VisionLite.faceAligner(context) {
    inputSize(256)
    threadCount(2)
    parallelism(2)
    cropMargin(0.25f)
}

val faceEmbedder = VisionLite.faceEmbedder(context) {
    preferredModelPreset(FaceEmbeddingModelPreset.MODERN_PREFERRED)
    threadCount(2)
    parallelism(2)
    normalizeOutput(true)
}

val faceAntiSpoofing = VisionLite.faceAntiSpoofing(context) {
    preferredModelPreset(FaceAntiSpoofingModelPreset.MODERN_PREFERRED)
    threshold(0.25f)
}

val faceQuality = VisionLite.faceQuality(context) {
    inputSize(256)
    minScore(0.48f)
    minConfidence(0.45f)
}

Detect and analyze a face

val faces = visionLite.faceDetector.detect(imageProxy)

val results = faces.map { face ->
    val crop = face.crop(imageProxy)
    val aligned = faceAligner.align(crop)
    val quality = faceQuality.assess(aligned.bitmap)
    val live = faceAntiSpoofing.check(aligned.bitmap)
    val embedding = if (live.live) faceEmbedder.embed(aligned.bitmap) else null

    FaceResult(
        face = face,
        live = live,
        modelQuality = quality,
        embedding = embedding,
    )
}

Match embeddings

val similarity = faceEmbedder.match(embeddingA, embeddingB)
val centroid = faceEmbedder.centroid(listOf(embeddingA, embeddingB, embeddingC))

---

Face modules

Module	What it does	Notes
Face detection	Finds faces and tracks them	ML Kit based, supports low-latency modes
Face alignment	Normalizes the face crop	Helps keep embedding and spoof checks stable
Face embedding	Generates face vectors	Supports match(...) and centroid(...)
Face anti-spoofing	Detects live vs spoofed faces	Keeps the live path safer
Face quality	Scores image quality	Useful for gating enrollment and capture

---

Module configs

Module	Main config	Common knobs
Face detection	FaceDetectorConfig	mode, trackingEnabled, minFaceSize, quality, maxConcurrentRequests
Face alignment	FaceAlignmentConfig	inputSize, modelAssetPath, threadCount, parallelism, cropMargin
Face embedding	FaceEmbedderConfig	preferredModelPreset, modelAssetPath, inputSize, outputSize, threadCount, parallelism, normalizeOutput
Face anti-spoofing	FaceAntiSpoofingConfig	preferredModelPreset, modelAssetPath, inputSize, threshold, threadCount, parallelism
Face quality	FaceQualityModelConfig	modelAssetPath, inputSize, minScore, minConfidence, threadCount, parallelism

Presets

• FaceEmbeddingModelPreset.MODERN_PREFERRED
• FaceEmbeddingModelPreset.FACE_NET_512

You can also set modelAssetPath(...) directly if you want to override the preset.

---

CameraX workflow

VisionLite is designed to work with ImageProxy directly so the hot path avoids unnecessary bitmap allocations.

Recommended flow:

1. Analyze preview frames with FaceDetector.
2. Pick the best face using coverage or tracking.
3. Crop the face with rotation correction.
4. Align the crop.
5. Run quality and anti-spoofing.
6. Generate the embedding only when the face is acceptable.
7. Store multiple embeddings per person and compare against the centroid.

---

Embedding strategy

VisionLite supports storing multiple embeddings for the same person and matching against a centroid.

Recommended enrollment flow:

1. Capture a live face.
2. Save the embedding.
3. Repeat up to a small number of samples.
4. Build the centroid from the stored embeddings.
5. Compare future faces against the centroid instead of a single sample.

This improves stability and reduces one-off false matches.

---

Good fit

• face enrollment
• face recognition
• camera preview overlays
• kiosk or on-device identity flows
• low-latency Android vision pipelines

---

Status

VisionLite is under active development.

The API is intentionally small and face-focused so the public surface stays easy to use and fast to review.

---

Contributing

See CONTRIBUTING.md.

Code of conduct

See CODE_OF_CONDUCT.md.

Security

See SECURITY.md.

License

See LICENSE.