Lumira：具备情绪、记忆与创作自主性的自主 AI 艺术家

An autonomous AI artist with personality, moods, memory, and creative independence. A lightweight "gallery + cloud-generation" build (Dockerfile.gallery, no local ML/GPU, generation via Replicate/Magica) deploys to Railway; see docs/DEPLOYMENT.md.

✨ What Makes Lumira Different

Lumira isn't a tool - she's an artist you collaborate with:

• Has Moods: 10 emotional states that influence her art
• Remembers: Episodic + semantic memory learns what works
• Reflects: Journals her thoughts after each creation (visible thinking with ReAct pattern)
• Chooses: Makes autonomous decisions about what to paint
• Evolves: Artistic style develops through experience and adaptive learning
• Ethical AI: Built-in bias mitigation to ensure fair and diverse output
• Production-Ready: 10x faster generation, PWA support, real-time monitoring

🚀 Performance Highlights

• ⚡ 10x Faster: First generation in 3-5s (was 35s) with model pool preloading
• 🎯 Smart Curation: GPU-parallel batch processing (3x faster quality scoring)
• 🧠 Adaptive Learning: Multi-armed bandit algorithm learns from user feedback
• 📱 Progressive Web App: Install as native app, works offline
• 🔴 Real-time Preview: WebSocket streaming shows generation progress
• 📊 Production Monitoring: Prometheus metrics + health checks for Kubernetes

Quick Start

# Clone and setup
cd lumira
python -m venv venv
source venv/bin/activate
pip install -r requirements-full.txt
pip install -e .

# Configure
cp config/config.example.yaml config/config.yaml

# Let Lumira create (she chooses based on mood)
python -m ai_artist.main

# Or suggest a theme
python -m ai_artist.main --theme "twilight dreams"

# Web gallery
python -m ai_artist.web.app

If you only want the web gallery (no ML generation), use:

pip install -r requirements.txt
pip install -e .

Project Structure

lumira/
├── src/ai_artist/
│   ├── personality/    # Moods, memory, identity
│   ├── core/           # Image generation
│   ├── curation/       # CLIP-based quality scoring
│   ├── scheduling/     # Autonomous creation
│   └── web/            # FastAPI gallery
├── config/             # Configuration
├── gallery/            # Generated artwork
└── docs/               # Documentation

📚 Documentation

📊 Current Status

See CLAUDE.md for the maintained current-state summary (stack, commands, gotchas). Highlights:

• 10-mood personality system with dynamic state
• 3-layer memory (episodic/semantic/working)
• Visible thinking process (ReAct pattern - observe/reflect/decide/express/create)
• Adaptive learning from user feedback (multi-armed bandit)
• Critique system for iterative improvement

Performance & Scale:

• Model pool with pre-warming (10x faster first generation)
• GPU-parallel batch curation (3x faster quality scoring)
• Kubernetes-ready health checks (liveness/readiness)
• Prometheus metrics (15+ production metrics)

User Experience:

• Progressive Web App (PWA) with offline support
• Real-time generation preview via WebSocket
• Modern, accessible dark UI with animations
• Autonomous scheduling
• FastAPI web gallery with WebSocket
• Service worker caching (3-tier strategy)

Quality & Safety:

• CLIP-based quality curation
• Ensemble curation framework
• Bias mitigation in prompt generation
• Intelligent prompt filtering
• Advanced prompt features (Matrix, Emphasis, Style Presets)
• Comprehensive test coverage (400+ tests)

See docs/ROADMAP.md for what's actively planned, and docs/ADVANCED_PROMPTS.md for prompt utilities.

🛠️ Tech Stack

Core:

• Stable Diffusion XL (diffusers + xFormers + SDPA)
• LoRA for style training
• PyTorch with torch.compile optimization

Backend:

• FastAPI + WebSocket for real-time updates
• SQLite + SQLAlchemy
• CLIP for quality curation
• Multi-armed bandit for adaptive learning

Frontend:

• Progressive Web App (PWA)
• Service Worker with 3-tier caching
• Modern responsive UI with WebSocket streaming

Production:

• Prometheus metrics
• Kubernetes health checks
• Sentry error tracking
• Comprehensive logging (structlog)

License

MIT License