---

✨ Showcase

Original	🪄 sanana Magic

---

🚀 One-Liner Start

# Install once
pip install sanana
sanana photo.jpg

# Or run without installing (requires uv)
uvx sanana photo.jpg

The model (~1.2GB) will be downloaded automatically on first run.

---

🔥 Highlights

🎨	Cartoon Style Transfer	Photos → Ghibli/anime illustrations
🏠	Fully Local	No internet, no cloud upload, absolute privacy
🐍	Python Native	pybind11 C++ extension, inference inside Python process
💻	CPU-only	MNN high-performance engine, runs on ordinary machines
⚡	10 Steps	Knowledge-distilled model, high quality with few steps
📦	Truly Ready-to-Use	pip install handles all dependencies automatically

---

📦 Installation

Option 1: pip (Recommended)

pip install sanana

This automatically clones and builds MNN (requires git, cmake, C++ compiler).

Option 2: From Source

git clone https://github.com/vra/sanana.git
cd sanana
uv pip install -e ".[dev]"

---

🎮 Usage

CLI

# Minimal
sanana photo.jpg

# Custom output
sanana photo.jpg -o output.jpg

# Custom prompt
sanana photo.jpg --prompt "Ghibli style"

# Full args
sanana photo.jpg \
  -o out.jpg \
  --prompt "Convert to Ghibli style" \
  --steps 10 \
  --seed 42 \
  --width 512 \
  --height 512

Argument	Description	Default
input	Input image path	Required
-o, --output	Output path	input_stem_cartoon.jpg
--prompt	Style prompt	Convert to a cartoon-style illustration
--steps	Inference steps	10
--seed	Random seed	42
--width/--height	Output size	512x512

Python API

from sanana import edit_image

# One-shot
edit_image("./photo.jpg", steps=10, seed=42)

from sanana import SanaEditor

# Advanced: reuse model for batch processing
editor = SanaEditor(backend="cpu", memory_mode=2)
editor.load()

for img in ["a.jpg", "b.jpg", "c.jpg"]:
    editor.run(
        input_image=img,
        output_image=img.replace(".jpg", "_cartoon.jpg"),
        steps=10,
    )

---

⚙️ Architecture

┌────────────────────────────────────────────┐
│  🐍 Python Layer                            │
│     sanana (CLI / API)                      │
│            ↓ import                         │
│     sanana._sana_diffusion                  │
│     (pybind11 C++ Extension)                │
└────────────────────┬───────────────────────┘
                     │
┌────────────────────▼───────────────────────┐
│  🔧 C++ Layer (MNN)                         │
│     SanaLlm ──→ SanaDiffusion ──→ Image     │
│     (Text Encode) (Flow Matching Denoise)   │
└────────────────────────────────────────────┘

• Text Encoding: Qwen3-0.6B LLM processes the prompt
• Feature Bridging: Connector → Projector maps to DiT space
• VAE Encode: Input image → Latent (img2img)
• DiT Denoise: Euler sampling + Flow Shift, 10 steps
• VAE Decode: Latent → Final image

---

🖥️ Performance

Apple MacBook Pro M3 (CPU, 512×512, 10 steps):

Stage	Time
LLM Text Encoding	~1.5 s
VAE Encode	~3 s
DiT Denoise (10 steps)	~22 s
VAE Decode	~5 s
Total	~32 s

---

🧩 Requirements

Platform	Status	Notes
macOS (Apple Silicon / Intel)	✅ Supported	Best performance on M1/M2/M3
Linux (x86_64 / ARM64)	✅ Supported	Tested on Ubuntu 22.04+
Windows (x86_64)	⚠️ Experimental	Requires Visual Studio + OpenCV, CI unstable

• Python: 3.10+
• Compiler: Clang/GCC/MSVC (C++17)
• Tools: git, cmake

---

❓ FAQ

Q: MNN not found during install?
A: setup.py auto-clones and builds MNN. Just ensure git, cmake, and a C++ compiler are installed.

Q: First run is slow?
A: The model (~1.2GB) is downloaded on first use. Subsequent runs are fast.

Q: GPU support?
A: CPU by default. Use backend="metal" for Metal (macOS), requires MNN built with GPU flags.

Q: Commercial use?
A: Code is Apache-2.0. Model files follow the original publisher's license.

---

🤝 Contributing

Issues and PRs welcome! See CONTRIBUTING.md.

📜 Changelog

See CHANGELOG.md.

📄 License

Code: Apache-2.0.
Model: MNN-Sana-Edit-V2 belongs to its original publisher.

---