AI Image Generation Integration for Mobile Apps

Generating images via Stable Diffusion, DALL·E 3, or Midjourney API – the bottleneck is not the algorithm but UX expectations and resource management. A cloud model request takes 5–30 seconds; on-device generation on mobile takes 10–60 seconds depending on model and device. Throughout that time, the user must understand what is happening. We have implemented similar mechanisms in dozens of projects and know how to avoid typical problems: from thermal throttling to content policy blocks. Contact us for a preliminary assessment of your project.

How to Integrate AI Image Generation into a Mobile App

Cloud Generation: DALL·E 3 and Stable Diffusion API

OpenAI Images API (POST /v1/images/generations) is the simplest path. The request returns an image URL or base64. Response time is 8–20 seconds for 1024×1024.

struct ImageGenerationRequest: Encodable {
    let model: String          // "dall-e-3"
    let prompt: String
    let n: Int                 // 1 (dall-e-3 does not support > 1)
    let size: String           // "1024x1024"
    let quality: String        // "standard" or "hd"
    let responseFormat: String // "url" or "b64_json"

    enum CodingKeys: String, CodingKey {
        case model, prompt, n, size, quality
        case responseFormat = "response_format"
    }
}

Replicate API provides access to Stable Diffusion XL, FLUX, and other open-source models. It uses an async model: the first request returns a prediction ID, then you need polling or a webhook. On mobile client, polling every 2 seconds with exponential backoff on errors:

suspend fun pollPrediction(predictionId: String): String {
    var delay = 2000L
    repeat(15) {
        delay(delay)
        val result = api.getPrediction(predictionId)
        if (result.status == "succeeded") return result.output.first()
        if (result.status == "failed") throw GenerationException(result.error)
        delay = minOf(delay * 1.5, 8000L).toLong()
    }
    throw TimeoutException("Generation timed out")
}

On-device Generation via Core ML

Apple ML Research released Stable Diffusion for Apple Silicon. On iPhone 15 Pro / M-series iPad – about 20 seconds for 512×512, 20 steps. On iPhone 12 – 60–90 seconds. The model weighs 2–6 GB depending on quantization.

import StableDiffusion

let pipeline = try StableDiffusionPipeline(
    resourcesAt: modelDirectory,
    controlNet: [],
    configuration: .init()
)
pipeline.loadResources()

var config = StableDiffusionPipeline.Configuration(prompt: userPrompt)
config.stepCount = 20
config.guidanceScale = 7.5
config.seed = UInt32.random(in: 0...UInt32.max)

let images = try pipeline.generateImages(configuration: config) { progress in
    DispatchQueue.main.async {
        self.generationProgress = Double(progress.step) / Double(progress.stepCount)
    }
    return true // continue generation
}

Thermal throttling is a real problem. After 3–4 consecutive generations, the iPhone drops performance. Solution: pause between generations, monitor ProcessInfo.thermalState, and warn the user.

On Android, on-device Stable Diffusion works via MediaPipe with LlmInferenceSession or directly through ONNX Runtime with GPU delegate. Support is significantly worse than on Apple Silicon – we recommend a cloud-first approach for Android.

What to Choose: Cloud or On-device Generation?

UX During Generation

A progress bar with a real value (not a spinner) is critical for long operations. Stable Diffusion returns progress.step – use it. Show intermediate previews (latent-preview) starting from step 5 – this keeps user attention.

Cancel generation: cloud request can be cancelled via URLSessionTask.cancel() or Replicate API POST /predictions/{id}/cancel. On-device – via a shouldContinue flag in the progress callback.

Save to gallery: PHPhotoLibrary.requestAuthorization(for: .addOnly) on iOS. WRITE_EXTERNAL_STORAGE permission (up to Android 9) or MediaStore.Images API. Request permission only at the moment of first save, not when opening the generation screen.

How to Manage Prompts and Avoid Content Policy Errors?

Content policy violations – DALL·E 3 rejects prompts with violence, NSFW, celebrity content. This requires prompt validation before submission (OpenAI Moderation API) and a clear error message. Do not show a system message “Your request was rejected” – explain what exactly is not allowed.

Device memory: on-device Stable Diffusion requires 4–6 GB RAM at peak. os_proc_available_memory() on iOS gives insight into available memory – if less than 1 GB is free, better to fall back to cloud.

Comparison of Integration Approaches

Work Process

Architecture selection → API integration → Generation UX → Error handling → Testing → Documentation. At each stage we provide intermediate results and agree on decisions with you. Order a turnkey integration and get a ready-made solution within agreed deadlines.

What Is Included in the Work

• Integration documentation (API specification, data schemas)
• Access to the code repository and CI/CD
• Training of the client's team (2–3 hours)
• Support during deployment to stores (App Store, Google Play)
• 3-month warranty on identified bugs

Timeline Estimates

Cloud generation with basic UI – 4–6 days. On-device Stable Diffusion with latent-preview and thermal management – 2–3 weeks. The exact estimate depends on the complexity of your project – contact us for a calculation.

Our team has 5+ years of experience in mobile development and has implemented more than 50 projects with AI generation. We guarantee quality at all stages – from prototype to production. Get a consultation today.

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.