Forge

Local-first AI image studio for Mac. Generate images with state-of-the-art diffusion models running entirely on your machine — no cloud, no API keys, no telemetry.

Features

• Text-to-image & image-to-image generation with real-time streaming preview
• 5 quantized models optimized for Apple Silicon (8GB–32GB RAM)
• Gallery with SQLite-backed search, metadata, remix, and export
• Model manager with one-click downloads, RAM detection, and recommendations
• First-run onboarding that guides you through your first model download
• 100% local — your images and prompts never leave your machine

Architecture

forge/
├── app/                    # Shared React UI package
│   └── src/
│       ├── components/     # Studio, Gallery, ModelManager, Onboarding
│       ├── hooks/          # useGeneration, useBackendHealth
│       ├── stores/         # Zustand state (server, generation)
│       ├── lib/api/        # API client + TypeScript types
│       └── pages/          # Route pages
├── tauri/                  # Tauri v2 desktop shell
│   └── src-tauri/          # Rust sidecar management
├── backend/                # Python FastAPI backend
│   └── backend/
│       ├── models/         # ModelManager, inference engine
│       ├── routes/         # REST + WebSocket endpoints
│       ├── database.py     # SQLite with FTS5 search
│       └── config.py       # Model registry
└── scripts/                # Build & dev helper scripts

Prerequisites

• macOS 13+ (Ventura or later) with Apple Silicon recommended
• Python 3.11+
• Node.js 20+
• Rust (latest stable) — install via rustup

Quick Start (Development)

1. Install dependencies

# Frontend dependencies
npm install

# Backend dependencies
cd backend
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
cd ..

2. Create the dev sidecar placeholder

Tauri requires the sidecar binary to exist at compile time, even in dev mode:

npm run setup:dev

3. Start the backend

In one terminal:

cd backend
source .venv/bin/activate
python -m backend.main

The backend runs at http://127.0.0.1:8188. You can verify with:

curl http://127.0.0.1:8188/api/v1/health

4. Start the frontend

In another terminal:

npm run dev

This starts the Vite dev server at http://localhost:5173 inside the Tauri window.

Browser-only mode

If you don't have Rust/Tauri installed yet, you can run just the Vite dev server directly:

cd tauri && npx vite

Then open http://localhost:5173 in your browser. The UI will connect to the backend at port 8188.

Supported Models

Model	RAM	Speed	Type
FLUX.2-klein-4B (4bit SDNQ)	8 GB+	Fast	txt2img + img2img
FLUX.2-klein-9B (4bit SDNQ)	12 GB+	Medium	txt2img + img2img
FLUX.2-klein-4B (Int8)	16 GB+	Medium	txt2img + img2img
Z-Image Turbo (Quantized)	8 GB+	Very fast	txt2img
Z-Image Turbo (Full)	24 GB+	Very fast	txt2img

Models are downloaded on-demand through the Model Manager or during onboarding.

Building for Distribution

# Build the backend into a standalone binary
npm run build:server

# Build the full Tauri app (includes backend sidecar)
npm run build

The .dmg / .app bundle will be in tauri/src-tauri/target/release/bundle/.

API Endpoints

The backend exposes these endpoints at http://127.0.0.1:8188/api/v1:

Endpoint	Method	Description
/health	GET	Backend health check
/system	GET	System info (RAM, platform, recommended model)
/models	GET	List all models with download/load status
/models/{id}/download	POST	Start model download
/models/{id}/download/progress	GET	Download progress
/models/{id}/load	POST	Load model into memory
/models/{id}/unload	POST	Unload model from memory
/models/{id}	DELETE	Delete model from disk
/generate	POST	Start image generation
/generate/stream	WebSocket	Streaming generation with live preview
/images	GET	List gallery images (with search)
/images/{id}	GET	Get image metadata
/images/{id}	DELETE	Delete image

Data Storage

All data is stored locally:

Platform	Location
macOS	~/Library/Application Support/Forge/
Linux	~/.local/share/forge/

Override with the FORGE_DATA_DIR environment variable.

Tech Stack

• Frontend: React 19, TypeScript, Tailwind CSS 4, Zustand, TanStack Query, Vite 6
• Desktop: Tauri v2 (Rust)
• Backend: Python, FastAPI, PyTorch, Diffusers, SDNQ/Quanto quantization
• Storage: SQLite with FTS5 full-text search
• Distribution: PyInstaller (backend binary), Tauri bundler (.dmg/.app)

License

Private — not yet open source.