Lumeo

Source: private Category: Project tour

Lumeo — a native iOS app that wraps Replicate’s image generation API behind a SwiftUI interface, with GPT-4o mini translating plain-language descriptions into model-ready prompts.

What it is

An iOS app, not a service. Users describe what they want in conversational language; GPT-4o mini converts that to an optimized prompt; Replicate generates the image; the app displays and stores results. All history lives locally.

Intent: make AI image generation usable by non-technical people. Raw prompt engineering (negative prompts, guidance scales, sampler choice) is hidden behind sliders and dropdowns.

Stack

Layer	Technology
Platform	iOS 17+, SwiftUI
Graphics	Metal shaders for mesh-gradient animations
API — images	Replicate prediction API
API — prompts	OpenAI GPT-4o mini
Persistence	UserDefaults (payloads + responses)
Design	Light/dark mode aware, compositional gallery layout

Flow

1. User types or speaks a description (“a cyberpunk cat at sunset”)
2. GPT-4o mini transforms that into an optimized prompt using tokenization-aware techniques
3. SwiftUI controls let the user tune guidance scale, steps, LoRA, aspect ratio, output quality
4. App submits a Replicate prediction, polls for completion, displays the result
5. Payload + response persist locally for recall

Non-obvious design choices

• UserDefaults over CoreData. Payloads are small and flat. UserDefaults is enough; CoreData overhead wasn’t justified.
• Polling, not webhooks. Replicate supports both; polling keeps the backend-free architecture intact. Webhooks would need a server the app can’t assume it has.
• Mesh gradient animation is Metal, not SwiftUI’s built-in. Needed more control over the shader; also a portfolio piece.
• Prompt assistant is deliberately opinionated. GPT-4o mini is instructed to produce prompts that work well on the specific Replicate models Lumeo supports, not generic ones.

Gotchas

• Replicate model IDs drift. Pinning a specific version hash means old users keep working; using latest means new users get the best model but old deployments can break. Pick an upgrade strategy.
• API keys in an iOS app are a problem. Embedded keys can be extracted from the binary. For personal use: fine. For distribution: back with a tiny proxy server that holds the real keys.
• Image results don’t persist remotely. Replicate’s result URLs expire. Lumeo caches images locally on first fetch; if it misses that window, history is orphaned. Cache aggressively.
• GPT-4o mini can hallucinate unsupported parameters. Validate the prompt assistant’s output against the Replicate model’s accepted fields before submitting.