AI Image Upscaling in Mobile Apps

AI Image Upscaling in Mobile Apps: On-Device and Cloud Solutions

You take a photo with your phone camera — 12 MP, but after cropping you're left with 600×600 pixels. Need 4× upscaling? Bicubic interpolation blurs everything. Neural networks restore textures of skin, fur, fabric. We implement AI upscaling three ways: on-device via Core ML / ONNX, cloud API, or hybrid. With over 5 years of experience and more than 30 AI projects, we guarantee quality results. On-device processing can save up to 80% on server infrastructure costs compared to cloud solutions. Contact us for a consultation on choosing the right approach.

import CoreML
import Vision

class ImageUpscaler {
    private let model: VNCoreMLModel

    init() throws {
        let config = MLModelConfiguration()
        config.computeUnits = .cpuAndNeuralEngine // Uses Neural Engine
        let coreMLModel = try RealESRGAN(configuration: config)
        model = try VNCoreMLModel(for: coreMLModel.model)
    }

    func upscale(_ image: UIImage) async throws -> UIImage {
        guard let cgImage = image.cgImage else { throw UpscaleError.invalidInput }

        return try await withCheckedThrowingContinuation { continuation in
            let request = VNCoreMLRequest(model: model) { request, error in
                if let error = error {
                    continuation.resume(throwing: error)
                    return
                }
                guard let results = request.results as? [VNPixelBufferObservation],
                      let outputBuffer = results.first?.pixelBuffer else {
                    continuation.resume(throwing: UpscaleError.noOutput)
                    return
                }
                let ciImage = CIImage(cvPixelBuffer: outputBuffer)
                let context = CIContext()
                guard let outputCG = context.createCGImage(ciImage, from: ciImage.extent) else {
                    continuation.resume(throwing: UpscaleError.conversionFailed)
                    return
                }
                continuation.resume(returning: UIImage(cgImage: outputCG))
            }
            request.imageCropAndScaleOption = .scaleFit

            let handler = VNImageRequestHandler(cgImage: cgImage)
            try? handler.perform([request])
        }
    }
}

Real-ESRGAN is the highest quality model for ×4 upscaling. Core ML-converted versions exist. Limitation: Real-ESRGAN expects a fixed input tile size (usually 256×256 or 512×512). For larger images, we slice into tiles with overlap (16–32 pixels) to avoid seams. (Real-ESRGAN: Xintao Wang et al., GitHub repository)

func upscaleTiled(_ image: UIImage, tileSize: Int = 256, overlap: Int = 16) async throws -> UIImage {
    let tiles = splitIntoTiles(image: image, tileSize: tileSize, overlap: overlap)
    let upscaledTiles = try await withThrowingTaskGroup(of: (Int, Int, UIImage).self) { group in
        for (row, col, tile) in tiles {
            group.addTask {
                let upscaled = try await self.upscale(tile)
                return (row, col, upscaled)
            }
        }
        var results: [(Int, Int, UIImage)] = []
        for try await result in group { results.append(result) }
        return results
    }
    return mergeTiles(upscaledTiles, originalSize: image.size, scaleFactor: 4, overlap: overlap)
}

On iPhone 15 Pro with Neural Engine: 512×512 → 2048×2048 takes ~800 ms. 1024×1024 split into tiles takes 2–4 seconds.

On-Device: ONNX Runtime on Android

Real-ESRGAN in ONNX format (~15 MB for the small model):

class OnnxUpscaler(context: Context) {
    private val session: OrtSession

    init {
        val env = OrtEnvironment.getEnvironment()
        val options = OrtSession.SessionOptions().apply {
            addNnapi() // Uses NNAPI for acceleration
        }
        val modelBytes = context.assets.open("realesrgan_x4.onnx").readBytes()
        session = env.createSession(modelBytes, options)
    }

    fun upscale(bitmap: Bitmap): Bitmap {
        // Convert Bitmap to float tensor [1, 3, H, W], normalize to [0, 1]
        val inputTensor = bitmapToTensor(bitmap)
        val inputName = session.inputNames.first()!!
        val output = session.run(mapOf(inputName to inputTensor))
        val outputTensor = output[0].value as Array<*>
        // Convert tensor back to Bitmap
        return tensorToBitmap(outputTensor, bitmap.width * 4, bitmap.height * 4)
    }
}

NNAPI on modern Android devices delivers 2–4× speedup over CPU. On Snapdragon 8 Gen 2 — 512×512 in ~1.2 seconds.

Cloud APIs

When on-device is too slow or you need a higher scaling factor (×8, ×16):

Replicate — Real-ESRGAN:

let body: [String: Any] = [
    "version": "...", // real-esrgan model hash
    "input": [
        "image": "data:image/jpeg;base64,\(base64Image)",
        "scale": 4,
        "face_enhance": true // GFPGAN for face enhancement
    ]
]

face_enhance: true runs GFPGAN on top of Real-ESRGAN — important for portraits to avoid artifacts.

Stability AI Upscale API:

val requestBody = MultipartBody.Builder()
    .setType(MultipartBody.FORM)
    .addFormDataPart("image", "photo.jpg", imageFile.asRequestBody("image/jpeg".toMediaType()))
    .addFormDataPart("output_format", "png")
    .build()

Stability AI returns PNG with ×4 upscaling via Creative Upscaler (SD-based, adds details) or Conservative Upscaler (fewer changes to original).

How to Integrate AI Upscaling: Step-by-Step

1. Model selection — Real-ESRGAN for on-device, cloud API for maximum quality.
2. Conversion to required format — Core ML for iOS, ONNX for Android.
3. Implement tiling — split into tiles with overlap, parallel processing.
4. UI integration — progress indicator, before/after slider.
5. Testing — on different devices, with various source images.

Approach Selection

Choosing Between On-Device and Cloud Upscaling

On-device is fast and free but limited by device performance and model size. Cloud APIs offer higher quality and scaling factors but add latency and cost. If your app targets offline scenarios or saves bandwidth, choose on-device. For maximum quality in portraits or old photos, cloud is justified. We implement a hybrid approach that automatically selects the best path.

Comparison: On-Device vs Cloud Upscaling

Why Real-ESRGAN Beats Standard Algorithms

Real-ESRGAN delivers 2–3× better quality than bicubic upscaling: it restores textures, noise reduction, sharpness. Combined with GFPGAN for faces, artifacts are virtually eliminated. This is proven in extensive testing and confirmed by the original paper. We use this model in most projects, including products with thousands of users.

UX: Progress and Comparison

On-device upscaling should show tiling progress: "Processing 3 of 12 tiles." The user doesn't experience freezes.

After completion, an interactive before/after slider (MagnificationGesture on iOS, custom touch view on Android) is standard UI for any photo enhancement tool.

What's Included

• Requirements analysis and optimal approach selection (on-device / cloud / hybrid)
• Development of native upscaling module (Swift / Kotlin)
• Integration of chosen model (Core ML, ONNX, cloud API)
• Tiling with overlap and progress indication
• Before/after slider
• Testing on real devices with different specifications
• Documentation and source code delivery
• One month of post-launch support

Estimated Timeline

On-device upscaling with tiling and progress indicator — 5–8 days. Cloud upscaling + face enhancement + before/after slider + gallery save — 8–12 days. Hybrid with quality evaluation and path selection — additional 3–5 days.

Contact us for a project assessment: get a consultation on approach selection and approximate cost.

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.