Forge — Local Image Generation Platform

Run Stable Diffusion on your own hardware. No cloud, no subscriptions, no data leaving your machine.

Forge is an open-source image generation platform that runs entirely on your local GPU. It provides a clean, modern web interface for creating images from text prompts — without requiring you to wire up complex node-based workflows or wrestle with command-line scripts.

Why Forge?

Existing local generation tools have a steep learning curve. ComfyUI is powerful but intimidating — new users face a blank canvas of nodes and wires before they can generate a single image. Automatic1111 works but shows its age with a Gradio interface that feels bolted together. Neither was designed as a polished product.

Forge takes a different approach:

• Zero-friction start — Type a prompt, click Generate. That's it. No nodes, no workflows, no YAML editing to get started.
• Fully local — Your prompts, images, and models never leave your machine. No API keys, no cloud accounts, no usage limits.
• Pluggable backends — Use HuggingFace Diffusers directly, connect to an existing ComfyUI instance, or run ONNX models on AMD/CPU. Switch backends from the settings page without restarting.
• Built for consumer GPUs — Targets RTX 4070 (12GB VRAM) as baseline. Automatic fp16, attention slicing, VAE tiling, and CPU offload keep memory usage in check.
• Real-time feedback — WebSocket-driven progress updates show you step-by-step previews as your image generates. No staring at a spinner wondering if it crashed.
• Modern stack — React 19, FastAPI, async everywhere. Fast to load, fast to develop on, easy to extend.

Quick Start

Prerequisites

• Python 3.11+
• Node.js 20+
• NVIDIA GPU with CUDA (recommended: RTX 4070 12GB+)

Setup

cd forge

# Backend
cd backend
python -m venv .venv
.venv/Scripts/pip install -e ".[dev,diffusers]"  # Windows
# source .venv/bin/activate && pip install -e ".[dev,diffusers]"  # Linux/Mac
cd ..

# Frontend
cd frontend
npm install
cd ..

# Configuration
cp forge.example.yaml forge.yaml
# Edit forge.yaml to set your model paths

Run

# Terminal 1 — Backend
cd backend && .venv/Scripts/python -m uvicorn forge.main:app --reload --port 7860

# Terminal 2 — Frontend
cd frontend && npm run dev

Open http://localhost:5173 in your browser.

No GPU? Forge ships with a demo backend that generates procedural images using Pillow. It activates automatically when PyTorch isn't installed, so you can explore the full UI without a GPU.

With Tilt (recommended for development)

pip install tilt
tilt up --port=10352

Starts backend (localhost:7860), frontend (localhost:5173), and Tilt dashboard (localhost:10352) with auto-reload.

Models

Place .safetensors model files in ~/.forge/models/checkpoints/. The directory is created automatically on first run.

Supported formats: .safetensors, .ckpt

Architecture

forge/
├── backend/           # Python FastAPI
│   └── forge/
│       ├── api/       # REST + WebSocket endpoints
│       ├── backends/  # Pluggable AI backends (diffusers, comfyui, onnx, demo)
│       ├── core/      # Job queue, GPU worker, event bus
│       ├── db/        # SQLAlchemy async + SQLite
│       ├── models/    # Model manager
│       ├── schemas/   # Pydantic request/response models
│       └── storage/   # Image + thumbnail storage
├── frontend/          # React + Vite + TypeScript
│   └── src/
│       ├── api/       # HTTP client + WebSocket manager
│       ├── stores/    # Zustand state management
│       ├── hooks/     # React hooks
│       ├── pages/     # Generation, Gallery, Models, Settings
│       └── components/# UI components
├── e2e/               # Playwright E2E tests (57 tests)
├── scripts/           # Dev tooling
├── Tiltfile           # Tilt development workflow
└── forge.example.yaml # Configuration template

Backend Design

• Pluggable backends — Abstract BaseBackend with concrete implementations for Diffusers, ComfyUI (API client), ONNX Runtime, and a GPU-free Demo. New backends register via decorator (@register_backend("name")).
• Serial GPU worker — Single consumer pulls jobs from an asyncio queue. One job at a time prevents OOM on consumer GPUs.
• Event bus — Job state changes broadcast to all connected WebSocket clients in real-time.
• SQLite + async SQLAlchemy — Zero-config persistence for job history, gallery metadata, and favorites.

Frontend Design

• Zustand for local UI state (generation params, canvas, queue)
• TanStack Query for server data (gallery, model list)
• Radix UI primitives for accessible components
• Tailwind CSS v4 with custom theme

API

Endpoint	Method	Description
/api/generate	POST	Submit a generation job
/api/jobs/{id}	GET	Get job status and results
/api/jobs/{id}/cancel	POST	Cancel a running job
/api/models	GET	List available models
/api/models/{id}/load	POST	Load a model into VRAM
/api/models/unload	POST	Unload current model
/api/gallery	GET	Browse generated images
/api/gallery/{id}/image	GET	Serve full-size image
/api/gallery/{id}/thumbnail	GET	Serve thumbnail
/api/gallery/{id}/favorite	PATCH	Toggle favorite
/api/settings	GET/PUT	Read/update configuration
/api/system/info	GET	System and GPU info
/api/system/health	GET	Health check
/ws/jobs	WebSocket	Real-time job progress

Testing

# Backend unit tests
cd backend && .venv/Scripts/python -m pytest tests/ -v

# Frontend type-check
cd frontend && npx tsc --noEmit

# E2E tests (starts backend + frontend automatically)
cd e2e && npx playwright test

Tech Stack

Backend: Python 3.11+, FastAPI, SQLAlchemy (async), SQLite, HuggingFace Diffusers, PyTorch

Frontend: React 19, Vite 6, TypeScript 5.7, Zustand 5, TanStack Query 5, Tailwind CSS v4, Radix UI, Lucide Icons

Testing: Playwright (E2E), Pytest (backend)

Roadmap

• Text-to-image generation
• Real-time progress with WebSocket
• Gallery with favorites
• Model management
• Demo backend (no GPU required)
• Image-to-image
• Inpainting / outpainting canvas
• LoRA support with weight sliders
• ControlNet (Canny, Depth, OpenPose)
• ComfyUI backend adapter
• ONNX backend (AMD / CPU)
• Batch generation
• CivitAI model downloads

License

MIT — Use it however you want.