IOS Demo Assets

iOS Demo Assets

Generate a batch of on-spec demo images with AI and get them Xcode-ready (valid Assets.xcassets bundles) and Simulator-ready (in the Photos library) in one pass. Runs on ModelRunner text-to-image via its public MCP. Every asset is just a prompt, so the whole library is reproducible and versionable in git — the same API key and request/poll pattern your app already uses, ModelRunner as dev infrastructure rather than just a user-facing backend.

Powered by ModelRunner — one API for image, video, audio, 3D, and text. Case study: https://modelrunner.ai/blog/generating-demo-assets-for-an-ios-app-with-ai

Prerequisites

1. The ModelRunner MCP must be connected — tools run_model, wait_for_request (and list_models to browse). Connect https://mcp.modelrunner.run/mcp (authorize with your ModelRunner account; keep a small balance at https://modelrunner.ai). Claude Code: claude mcp add --transport http modelrunner https://mcp.modelrunner.run/mcp.
2. Xcode + a booted iOS Simulator, and a shell for curl / xcrun.

Step 1 — Plan the set

Agree the categories, counts, and kebab-case asset names up front — they become .imageset folder names and SwiftUI Image("Group/name") keys. Example layout:

Assets.xcassets/
├── SampleRooms/        room-living-modern.imageset, room-kitchen-modern.imageset, …
├── SampleProducts/
│   ├── Interior/       product-sectional-sofa.imageset, product-floor-lamp.imageset, …
│   └── Exterior/       product-outdoor-sofa.imageset, …
└── Exterior/           exterior-garden-plain.imageset, exterior-balcony-empty.imageset, …

Step 2 — Generate (async, parallel)

Browse the live catalog with list_models({ category: "text-to-image" }) and confirm pricing with get_model — the catalog changes. Known-good picks: qwen/qwen-image (default; per-megapixel, with accurate in-image text), google/imagen4/fast (fast, cheap, photoreal), black-forest-labs/flux-2 (strong prompt adherence), plus newer volume/photoreal options such as tongyi-mai/z-image/turbo (tuned for cost-efficient bulk) or krea/krea-2-large (photoreal). For faces/hands where realism must be exact, a follow-up edit pass helps. For each asset:

1. run_model({ endpoint: "qwen/qwen-image", input: { prompt } }) → returns a requestId immediately (async queue).
2. wait_for_request({ requestId }) → the terminal output is a string[]; take output[0].
3. curl -L -o <path>.jpg "<output[0]>".

Submit the whole batch concurrently — each request resolves independently, so a 20–30 image library finishes in a few minutes rather than serially. Prompt recipes (photoreal scenes vs. white-background product shots) are in references/prompt-recipes.md.

Step 3 — Write into Assets.xcassets

For each downloaded image, build a valid image set:

python3 scripts/make_imageset.py <image.jpg> <Assets.xcassets/Group/name>

It creates <name>.imageset/ containing the image plus a Contents.json that Xcode's asset catalog compiler understands. The image is then addressable in SwiftUI as Image("Group/name").

Step 4 — Push into the Simulator's Photos library

The in-app photo picker (PhotosPicker) reads the Simulator Photos library, not Assets.xcassets. Get the booted device UDID and add the media:

xcrun simctl list devices booted          # copy the UDID
scripts/add_to_simulator.sh <UDID> <one-or-more .jpg files>

add_to_simulator.sh wraps xcrun simctl addmedia. The Photos library is per-simulator, so re-run this against each new simulator UDID — the xcassets copy stays the durable source of truth.

Reproducibility, cost, output

• Keep the prompts in git next to the app. Need a warmer palette or a localized set? Re-run the same batch — no stock-photo relicensing or re-shoot.
• Cost is per-megapixel on qwen/qwen-image (~$0.016/image) — a flat, known price for a full regeneration; confirm live with get_model.
• Report the asset names + ModelRunner request IDs + output URLs so the run is reproducible.
• Scale — in the case study this produced 29 assets across 4 categories in under 10 minutes of wall-clock (parallel generation), replacing 54 hand-picked stock PNGs and shrinking the bundle.