Foundation Models Skill

When to Load

Load this skill when:

• •Implementing AI-generated images for custom words
• •Working with ImagePlayground framework
• •Handling async image generation
• •Managing on-device AI memory/performance

Key Frameworks

ImagePlayground

For generating images from text prompts:

import ImagePlayground

FoundationModels (Optional)

For text generation with on-device LLM:

import FoundationModels

ImageCreator Usage

Basic Pattern

Task.detached(priority: .userInitiated) {
    do {
        let creator = try await ImageCreator()
        let concepts: [ImagePlaygroundConcept] = [.text(prompt)]
        let images = creator.images(for: concepts, style: .illustration, limit: 1)
        
        for try await image in images {
            let uiImage = UIImage(cgImage: image.cgImage)
            await MainActor.run {
                self.state.roundState?.generatedImage = uiImage
            }
            break // Only need first result
        }
    } catch {
        print("Generation failed: \(error)")
    }
}

Available Styles

• •.photo – Photorealistic
• •.illustration – Illustrated/artistic
• •.sketch – Pencil sketch style
• •.monochrome – Black and white
• •.fantasy – Fantastical/imaginative

Concepts

.text("a medieval castle")           // Text description
.extractedFrom(url)                  // Extract from image URL
.person(...)                         // Person-based (requires permissions)

Important Considerations

1. •Device Support: Requires A12 Bionic or later
2. •Memory: Model uses ~few hundred MB when active
3. •Time: Generation takes 1-3 seconds typically
4. •Off-Main: Always run in detached Task
5. •Content Safety: Framework has built-in content filtering
6. •Fallback: Handle errors gracefully (no image is acceptable)

UI Alternative

For user-controlled image creation (not recommended for game flow):

.imagePlaygroundSheet(isPresented: $showPicker, concept: .text(prompt)) { url in
    // Handle generated image URL
}

Memory Management

• •Create ImageCreator in local scope
• •Don't hold reference longer than needed
• •Clear generatedImage when round ends
• •Apple manages model unloading automatically

Critical Research Notes (Section 14)

Device Requirements

• •Requires A12 Bionic or later
• •Model is ~3GB quantized, few hundred MB in active memory
• •Available on iOS 26+ without special entitlements

Content Safety

• •Framework has built-in content filtering
• •May refuse certain prompts for safety
• •Handle gracefully - no image is acceptable fallback

Alternative: ImagePlaygroundSheet

For user-controlled image creation (NOT recommended for game flow):

.imagePlaygroundSheet(isPresented: $showPicker, concept: .text(prompt)) { url in
    // Handle generated image URL
}

This interrupts flow, so we use programmatic ImageCreator instead.

SystemLanguageModel (Optional)

Could use for generating random word from theme:

// Pseudo: "Give me one random word related to [theme]"
// But adds unpredictability - we skip this for MVP

Timing Considerations

• •Generation takes 1-3 seconds typically
• •Kick off during startGame action
• •Overlaps with role reveal pass-and-play time
• •Show ProgressView if not ready when needed