Generate images with AI — entirely on your Mac. No cloud, no API keys, no subscriptions.
LocalImg is a native macOS app that runs Stable Diffusion models locally on Apple Silicon using the incredible MLX framework. Type a prompt, hit generate, and watch your image come to life — all without a single byte leaving your machine. This is made for fun, MVP stage, curious on local image generation project.
- Text-to-Image Generation — Describe what you want, and LocalImg generates it right on your Mac.
- 100% Local & Private — Everything runs on-device. No internet required after the initial model download. Your prompts never leave your computer.
- Live Preview — Watch the image progressively denoise in real-time as it generates.
- Generation History — Every image you create is saved with its full parameters (prompt, seed, steps, guidance, dimensions) via SwiftData, so you can browse, revisit, and reproduce any past generation.
- Zoomable Image Preview — Pinch-to-zoom and pan across your generated images with smooth controls.
- Adjustable Parameters — Fine-tune steps, guidance scale, seed, dimensions, and output format (PNG/JPEG) to your liking.
- Multiple Models — Ships with support for SDXL Turbo, with an architecture that makes adding more models straightforward. (Support for Stable Diffusion 2.1 Base is still baking)
This project exists because of Apple's MLX framework — and honestly, it's kind of magic.
MLX is an array framework designed specifically for machine learning on Apple Silicon. It brings GPU-accelerated inference to the Mac in a way that feels native, efficient, and fast. The mlx-swift-examples repository provides a Swift-native Stable Diffusion implementation that this app wraps, meaning we get:
- Unified Memory — Apple Silicon shares memory between CPU and GPU, so there's no expensive data copying. The model just runs.
- Float16 & Quantization — MLX supports fp16 and quantized inference out of the box, keeping VRAM usage low even on machines with 8 GB of unified memory.
- Lazy Evaluation — Computations are only materialized when needed, which keeps the pipeline lean.
- Swift-Native — No Python bridge, no ONNX runtime, no Core ML conversion step. Pure Swift, directly calling into MLX's metal shaders.
Apple's investment in on-device ML is genuinely impressive, and MLX is the clearest expression of that vision. This project is a small love letter to that work. 💜
LocalImg is a pure Swift Package — no .xcodeproj, no .xcworkspace. You can build and run it with swift run from the terminal. This keeps the project lightweight, portable, and free from Xcode-specific cruft. The NSApplication.setActivationPolicy(.regular) call in the app entrypoint ensures it behaves as a proper foreground GUI app even when launched from the terminal.
- SwiftUI because it's the modern, declarative way to build macOS interfaces, and it pairs beautifully with the
@Observablepattern for reactive state management. - SwiftData for history persistence because it's zero-config, built into the platform, and perfect for the simple schema of generation records.
The ImageGenerationEngine protocol abstracts the generation backend. Today it's MLXDiffusionEngine, but this design means you could swap in a Core ML backend, a remote API, or anything else without touching the UI layer.
The engine auto-detects available system RAM and adjusts its strategy:
- < 8 GB: Quantized weights, aggressive memory limits, conservative cache.
- ≥ 8 GB: Full fp16 weights with a generous 256 MB cache for faster subsequent generations.
- macOS 14 (Sonoma) or later
- Apple Silicon (M1 / M2 / M3 / M4) — this will not run on Intel Macs
- ~40 GB of free disk space for the SDXL Turbo model
git clone https://github.com/scodes73/local-img.git
cd local-img
swift runOn first launch, the onboarding screen will guide you through downloading your chosen model. After that, you're fully offline-capable.
I highly recommend to download the model using huggingface-cli (to not get rate limiting and higher download speeds):
brew install huggingface-cli
hf download stabilityai/sdxl-turboThis is a Minimum Viable Product built for fun. 🎉
LocalImg was made as a passion project to explore what's possible with MLX on Apple Silicon. It is not production-grade software. You may encounter:
- Rough edges in the UI
- Occasional generation failures on unusual dimension/step combos
- Limited model selection (just two presets today)
- No batch generation, img2img, inpainting, or ControlNet (yet?)
The code is written to be clear and hackable over bulletproof. If you want to experiment with on-device image generation on your Mac, this is a fun starting point — not a finished product.
PRs, ideas, and vibes are welcome. 🤙
Sources/LocalImg/
├── LocalImgApp.swift # App entry point
├── MLX/ # MLX Stable Diffusion primitives
│ ├── Clip.swift # CLIP text encoder
│ ├── Configuration.swift # Model configurations
│ ├── Image.swift # Image utilities
│ ├── Load.swift # Weight loading
│ ├── Sampler.swift # Noise sampling
│ ├── StableDiffusion.swift # Pipeline orchestration
│ ├── Tokenizer.swift # Text tokenization
│ ├── UNet.swift # UNet denoiser
│ └── VAE.swift # Variational autoencoder
├── Models/
│ ├── AppSettings.swift # UserDefaults-backed settings
│ ├── GenerationParameters.swift # Prompt, steps, seed, etc.
│ ├── GenerationRecord.swift # SwiftData history model
│ └── ModelInfo.swift # Model metadata & presets
├── Services/
│ ├── HistoryManager.swift # SwiftData CRUD
│ ├── ImageGenerationEngine.swift # Engine protocol
│ ├── MLXDiffusionEngine.swift # MLX backend
│ ├── ModelManager.swift # Download & cache management
│ └── ThumbnailLoader.swift # Async thumbnail loading
└── Views/
├── ContentView.swift # Main layout
├── GenerationView.swift # Prompt input & image display
├── HistoryGalleryView.swift # History sidebar
├── HistoryThumbnailView.swift # History thumbnails
├── ImagePreviewView.swift # Zoomable image preview
├── LiquidGlassView.swift # Animated background effect
├── OnboardingView.swift # First-launch setup
├── SettingsView.swift # Preferences window
└── ZoomableImageView.swift # Pinch/scroll zoom
- Apple MLX — For making on-device ML on Apple Silicon genuinely delightful.
- mlx-swift-examples — The Swift Stable Diffusion implementation that powers this app.
- Hugging Face — For hosting the model weights and making them accessible.
- Stability AI — For the Stable Diffusion and SDXL model families.