Face Swap Pipeline: Detection, Swap, Moderation

Many teams face two main problems when implementing face swap: app rejection due to store policy violations and insufficient swap quality—artifacts, skin color mismatch, unnatural contours. Let's examine the architecture of a production-ready face swap pipeline, including on-device and API approaches, post-processing, content moderation, and App Store and Google Play syndicate requirements. Contact us to evaluate your project.

Why on-device face swap is rare?

Models like SimSwap (300 MB), FaceShifter (250 MB), GHOST (100 MB) require GPU. TFLite ports lose ~15% quality. MediaPipe Face Mesh provides 468 landmarks in real time — we use them for alignment. However, the swap itself requires neural network inference. On-device implementation is possible via Core ML on iPhone 14 Pro+: a distilled model processes a frame in 1–3 seconds. On Android — TFLite GPU delegate, but behavior varies (Adreno 730 vs PowerVR). Production apps often use APIs: InsightFace, Akool, DeepFaceLab. Akool face_swap endpoint returns results in 5–15 seconds.

The API approach is faster to implement and saves device resources, but on-device guarantees privacy and offline work. We offer both options depending on client tasks.

How is post-processing performed after face swap?

Even a good face swap produces artifacts. On iOS we apply CIFilter(name: "CIGaussianBlur") on the face mask and CIBlendWithMask for smooth transitions; for complex cases — Metal Performance Shaders. Skin color between face and neck may mismatch — colour transfer via LAB color space with CIColorCube filter. On Android we use OpenCV for GaussianBlur and inpaint. Infrastructure savings with on-device can be significant compared to API under high load.

Client-side processing pipeline

Preprocessing is needed before server submission:

// Android: face detection and cropping before submission
class FacePreprocessor(private val context: Context) {
    private val detector = FaceDetection.getClient(
        FaceDetectorOptions.Builder()
            .setPerformanceMode(FaceDetectorOptions.PERFORMANCE_MODE_ACCURATE)
            .setLandmarkMode(FaceDetectorOptions.LANDMARK_MODE_ALL)
            .build()
    )

    suspend fun extractFace(bitmap: Bitmap): FaceExtractionResult {
        val image = InputImage.fromBitmap(bitmap, 0)
        val faces = detector.process(image).await()
        if (faces.isEmpty()) throw FaceSwapError.NoFaceDetected
        if (faces.size > 1) throw FaceSwapError.MultipleFaces

        val face = faces.first()
        val bounds = face.boundingBox
        // Expand bounds by 40% for better context
        val expandedBounds = expandRect(bounds, 0.4f, bitmap.width, bitmap.height)
        val croppedBitmap = Bitmap.createBitmap(
            bitmap, expandedBounds.left, expandedBounds.top,
            expandedBounds.width(), expandedBounds.height()
        )
        return FaceExtractionResult(croppedBitmap, face.headEulerAngleY)
    }
}

Head rotation angle (headEulerAngleY) is important: if deviation >30°, swap quality drops sharply — the user should be warned.

Work process

1. Analytics: determine source (photo/video), choose approach (on-device/API), evaluate metrics.
2. Design: architecture development, model selection (Core ML/TFLite), UI design.
3. Implementation: integrate face detection (MediaPipe, ML Kit), preprocessing, request submission, post-processing.
4. Testing: check on different devices, A/B quality testing, stress tests.
5. Deployment: publish to App Store/Google Play, configure moderation, monitor.

What's included in the work

The full package includes:

• Architecture documentation and API integration specs (endpoint specifications, postman collection).
• API access (test keys, sandbox).
• Client team training (2-3 workshops on pipeline and moderation).
• Technical support during implementation and after release (1 month).
• Integration of watermark and disclaimer, Terms of Service setup per store requirements.

Legal restrictions and moderation

Face swap is subject to strict requirements under App Store Review Guidelines (Guideline 1.1 Objectionable Content). Apple rejects apps that:

• allow inserting another person's face without explicit consent;
• lack watermark or AI-generated labeling;
• could be used for deepfakes.

Mandatory minimum: watermark on result, explicit disclaimer in onboarding, Terms of Service prohibiting use of real faces without consent, content reporting system.

Content moderation on the server: before generation, run through Amazon Rekognition DetectModerationLabels or Google Cloud Vision Safe Search. If input photo is flagged — reject on backend, do not proceed to generation.

Storage and deletion

Face swap results should not be stored on the server longer than needed for client delivery. Standard practice: TTL 24-48 hours, then auto-delete from S3/GCS. Input photos — delete immediately after processing.

Timeline

Basic API integration with face detection and result display — 4-6 days. With blending post-processing, content moderation, watermarking, and store policy compliance — 3-4 weeks. Cost is calculated individually after auditing your project. We are a team of certified developers with 5+ years of experience in mobile development and AI, having completed 10+ face swap projects. Order a test integration or get a consultation — we'll evaluate your project for free.

Machine Learning in Mobile Apps: CoreML, TFLite, and On-Device Models

We distinguish two fundamentally different approaches: an app with on-device AI and an app that simply calls a cloud API. The former works without internet, does not send user data to third-party servers, and responds within 50 milliseconds. The latter depends on network latency and pricing plans. Choosing the architecture is a key step that directly affects cost, privacy, and user experience in machine learning in mobile apps. Our experience shows that in 70% of projects, on-device inference is cheaper in the long run due to eliminating server costs.

How to Choose Between CoreML and TFLite for On-Device Inference?

CoreML — Apple's native framework for running ML models on device. Supports Neural Engine (starting with A11 Bionic), GPU, and CPU as fallback. Models are converted to .mlmodel format via coremltools from PyTorch, ONNX, or TensorFlow. Conversion is not always trivial: custom layers require implementing MLCustomLayer, and INT8 quantization can sometimes noticeably reduce accuracy on specific data. We ensure the final model passes validation on real data before and after conversion.

TensorFlow Lite — cross-platform alternative for Android and Flutter. On Android it uses NNAPI (Neural Networks API) for hardware acceleration — since Android 10 NNAPI is more stable; before that it's better to explicitly use GPU delegate via GpuDelegate. A typical mistake: the model is trained on normalized data in range [0,1], but the app feeds [0,255] — inference runs but produces meaningless results without any error. We include an automatic input data validation module in the SDK.

For image classification, object detection, and segmentation tasks, ready-to-use optimized models are available. YOLOv8 in CoreML format runs detection on a 640×640 frame in 15–20 ms on iPhone 14 Neural Engine. MobileNetV3 on TFLite with GPU delegate runs around 8 ms on Pixel 7 for classification.

What If You Need Text Generation On-Device?

Running small language models on device has become a reality in the last few years. Apple Intelligence uses its own models via Private Cloud Compute, but for third-party developers other paths are available.

llama.cpp with Metal backend on iOS is a working approach for phi-3-mini (3.8B parameters, 4-bit quantization, ~2.3 GB). Inference: 15–25 tokens/second on iPhone 15 Pro. For integration in Swift, use the Swift Package llama.swift or a wrapper via C interface llama.h. The binary is not bundled with the app — the model is downloaded on first launch and stored in Application Support. Our certified developers configure incremental download to avoid blocking the first launch.

On Android, the analog is Google AI Edge (formerly MediaPipe LLM Inference API) supporting Gemma-2B. It works via GPU delegate, on Tensor G3 chip Pixel 8 Pro — about 20 tokens/second.

Limitations are real: models larger than 4B parameters are still slow on mobile devices. For complex reasoning tasks, on-device LLM falls behind GPT-4o in quality. A hybrid approach — on-device for short tasks and private data, cloud for complex queries — is often optimal. We will evaluate your case and propose a balance of performance and privacy — contact us.

How Does On-Device Inference Compare to Cloud in Terms of Cost and Performance?

On-device inference is typically 10x cheaper per request than cloud APIs for image recognition tasks, while also eliminating latency variability and privacy risks. The table below summarizes the trade-offs.

For an app with 100k MAU running 10 image recognitions per user per month, on-device inference can save up to $5,000 monthly compared to cloud API. Get a free consultation on your ML architecture today.

Integrating OpenAI API and Other Cloud Models

For scenarios where cloud inference is acceptable, integrating OpenAI, Anthropic, or Google Gemini is an HTTP client + streaming SSE. In Swift, AsyncThrowingStream is convenient for streaming responses. In Kotlin, use Flow.

Critically: API keys must never be stored in the app bundle. Even an obfuscated key can be extracted from the IPA in 10 minutes using strings or frida. Correct architecture: mobile app → your own backend → OpenAI API. The backend controls rate limiting, logs requests, and protects the key.

What Is Included in the Work (Deliverables)

• Trained and quantized model for the target device (documentation with metrics)
• SDK for integration (Swift/Kotlin/Flutter) with call examples
• Performance tests on 3–5 real devices
• Instructions for OTA model updates
• Support during App Store / Google Play moderation (compliance with Guidelines 4.2, 5.1)
• 2 weeks of technical support after release

Typical Project Pipeline

1. Task analysis — measure latency, privacy, size, supported devices.
2. Model prototyping — in Python, evaluate accuracy on target data.
3. Conversion and quantization — for CoreML/TFLite with validation.
4. Integration into the app — model wrapped in a service layer (easy to swap CoreML ↔ TFLite ↔ cloud).
5. Testing — on real devices, measure FPS, RAM, battery.
6. Deployment — via TestFlight / Firebase App Distribution, monitor metrics.

Timelines: integration of a ready CoreML/TFLite model — 1–2 weeks, development of a custom model with mobile optimization — from 6 weeks, on-device LLM chat with personalization — 4–8 weeks.

Why We Take on Complex Cases?

10+ years of experience in mobile development, 50+ implemented AI/ML solutions, guarantee of compatibility with current iOS and Android versions. All projects undergo code review and load testing. The cost includes preparation of moderation documentation and training of your team.

Contact us — we will help you choose the architecture and implement ML in your app turnkey. Order an audit of your existing solution — we will assess the potential for server cost savings free of charge. In some projects, savings can reach significant amounts per month.