Building Scevia: lessons from shipping a generative AI photo and video app on iOS

Restoring old photos and turning a single still into a short video sounds simple on a slide. Here is what it actually takes to ship it as a native iOS app.

Mustafa GüderFounder, GuderTechUpdated 4 min read
building-a-generative-ai-ios-app.md

Scevia is our AI creative studio for iPhone. It restores and colourises old photographs, turns a single still into a 5 to 15 second vertical video, places people into new scenes and relights photos. It ships in six languages. This is what building it taught us.

Keep the AI behind your own server

It is tempting to call an AI provider straight from the app. Do not. Your API key would ship inside the app binary, where anyone can extract it, and you would have no way to limit abuse.

Instead, the app talks to our backend, and the backend talks to the models. That gives us:

  • Keys that never leave the server
  • Per-user limits and credits enforced in one place
  • Freedom to switch or mix models without an app update
  • Logs to debug failed generations

We use Supabase for authentication, storage and server functions, which keeps the backend small and the bills predictable.

Design for waiting

Image restoration takes seconds; video generation can take much longer. A spinner is not enough. What works:

  1. Start the job on the server and return immediately
  2. Show progress and let people leave the screen
  3. Send a push notification when the result is ready
  4. Keep finished results in a library, so nothing is lost if the app is closed

Make the cost visible to yourself first

Every generation costs money. Before launch we modelled cost per action, then set credits and pricing so that a heavy user is still profitable. Free trials are generous enough to show the magic, and limited enough to survive abuse. If you are planning an AI product, include these running costs in your budget; our iOS app cost guide explains the rest.

Show the result before the paywall

The onboarding shows before-and-after examples and lets people try the core feature first. Users understand an AI app in one image, not in a paragraph of text.

Moderation is part of the product

Any app that generates images from user photos needs rules: what can be generated, which providers see the data and how long files are kept. In Scevia, uploads and results are tied to the account and removed when the account is deleted, and the privacy policy names every provider involved in plain language. Clear answers here also make App Review smoother.

Guest mode deserves the same care. People can try Scevia without an account, and a random identifier in the iOS Keychain makes sure the free welcome credits can only be claimed once per device, without linking it to anyone's identity.

Six languages from day one

Localizing the interface, the App Store listing and the onboarding examples opened markets we would never have reached in English alone. With String Catalogs in Xcode, adding a language is mostly translation work, not engineering. See our ASO checklist for the listing side.

Native where it counts

Scevia is built in Swift and SwiftUI. Smooth before-and-after sliders, video playback, photo library access and StoreKit 2 subscriptions all feel better native, and they are exactly the parts people touch most. Our StoreKit 2 guide shows how we handle subscriptions.

What we would tell a founder building an AI app

  • Start with one feature people will pay for, not ten demos
  • Put a server between your app and every model
  • Model your costs per user before setting prices
  • Design the waiting experience as carefully as the result
  • Write your privacy and moderation rules before App Review asks

If you want to add AI to an existing app or build a new one, talk to us. Scevia is the reference we build from.

Frequently asked questions

Can I call OpenAI or other AI APIs directly from my iOS app?

You can, but you should not: the key can be extracted from the app. Route requests through your own backend so you control keys, limits and costs.

How much does it cost to run an AI app?

It depends on the model and on usage. Image generation costs cents per request; video can cost much more. Model the cost per action before you set prices.

Does Apple accept AI-generated content apps?

Yes, as long as the app has clear content rules, filters harmful requests and explains how user data is handled.

Can AI features run on the device instead?

Some can. Apple's on-device models and Core ML handle tasks like classification and text. Large image and video models still run on servers.

ShareXLinkedInWhatsApp
Mustafa Güder

About the author

Mustafa Güder

Mustafa Güder is the founder of GuderTech and an iOS developer. After two years of study in Australia he came home to Adana and built Scevia, Paw & Home and Street Legends.

Studied in Australia, 2 years Adana, Türkiye

All posts