Mine StableDiffusion 🎨

The kotlin multiplatform Stable Diffusion client
Generate stunning AI art locally on Your devices

---

✨ What is Mine StableDiffusion?

Mine StableDiffusion is a native, offline-first AI art generation app that brings the power of Stable Diffusion models to your fingertips. Built with modern Kotlin Multiplatform technology and powered by the blazing-fast stable-diffusion.cpp engine, it delivers desktop-class performance on both Android/iOS and Desktop platforms.

🎯 Why Choose This App?

• 🚀 Native Performance - C++ backend with JNI bindings for maximum speed
• 🔒 Privacy First - 100% offline, all processing happens on your device
• 🎨 Modern UI - Beautiful Compose Multiplatform interface
• 📱 True Multiplatform - Shared codebase for Android & iOS & Desktop
• 🔧 Model Flexibility - Support for FLUX, SDXL, SD3, and many more
• ⚡ Hardware Accelerated - Vulkan 1.2+ (Android/Linux/Windows) & Metal (macOS)

---

📸 Screenshots

🤖 Android-1	Android-2	Android-3

💻 Desktop-1	💻 Desktop-2	Desktop-mac

---

🎲 Supported Models & Performance Tiers

Mine StableDiffusion supports a wide range of models. To help you choose the best model for your device, we've organized them by performance requirements:

Please try to ensure that the model is smaller than the VRAM of your device.

[!TIP] Start Small: We recommend starting with smaller models (e.g., SD-Turbo, SD 1.5) and gradually trying larger ones. This allows you to gauge your device's capabilities and identify performance bottlenecks effectively.

� Entry & Speed (Fastest, Minimal VRAM)

Ideal for older phones or integrated graphics. High speed, low memory usage.

• ✅ SD-Turbo - Extremely fast 1-step generation
• ✅ SD1.x / SD2.x/Illustrious
• ✅ SDXL-Turbo - Fast high 512x512
  • Test:Model
  • 

⚖️ Balanced Performance (Standard)

Good balance between quality and speed. Works well on most modern devices.

• ✅ 🎨 Chroma / Chroma1-Radiance - Vibrant color generation
• ✅ 🖼️ Z-Image - Advanced image synthesis
  • Test:Model+VAE+LLM
  • 

💎 Professional Quality (High Requirements)

Best for high-detail 1024x1024+ generation. Requires more VRAM and time.

• ✅ SDXL - Standard high-quality base model
• ✅ SD3 / SD3.5 - Stability AI's latest high-fidelity architecture
• ✅ 👁️ Ovis-Image - Vision-language model

🌌 Next-Gen Large Models (Flagship & High-End PC)

State-of-the-art models with massive parameter counts. Best for flagship phones or dedicated GPUs.

• ✅ FLUX.1-schnell / dev - Next-gen image quality
• ✅ FLUX.2-dev - Latest and most capable iteration

---

🌟 Key Features

Text-to-Image Generation

Generate stunning images from text descriptions with various models

Input: "A serene mountain landscape at sunset, digital art"
Output: High-quality AI-generated image

⚙️ Advanced Settings Guide

The Advanced Settings page provides fine-grained control over the inference engine. Below is a summary of each toggle and its impact:

Setting	Description	Effect When ON	Effect When OFF	Recommendation
Offload to CPU	Offloads model computations from GPU to CPU	Saves GPU/VRAM at the cost of slower generation speed.	All computation stays on GPU (faster but needs more VRAM).	Enable on low-VRAM devices.
Keep CLIP on CPU	Forces the CLIP text encoder to stay on CPU	Frees GPU memory for image generation; slightly slower prompt encoding.	CLIP runs on GPU (faster but uses more VRAM).	✅ Enabled by default on macOS to prevent potential crashes.
Keep VAE on CPU	Forces the VAE decoder to stay on CPU	Frees GPU memory; decoding step is slower.	VAE runs on GPU (faster final decode).	Enable if you encounter OOM errors during decode.
Enable MMAP	Memory-maps model weights from disk instead of loading them entirely into RAM	Lower initial RAM spike; the OS pages weights in on demand (more disk I/O).	Entire model is loaded into RAM upfront (higher peak RAM, lower disk I/O).	Disable if you experience slow generation on devices with slow storage.
Direct Convolution	Uses a direct convolution implementation in the diffusion model	Experimental performance boost on some hardware.	Standard im2col-based convolution is used.	Try enabling to see if it improves speed on your device; disable if quality degrades.

Model Weight Type (wtype) — Controls how model weights are stored in memory. Lower bit-depth reduces RAM usage but may degrade image quality.

Tip

K-variants (Q6_K, Q5_K, Q4_K, Q3_K, Q2_K) offer better quality at the same bit-depth compared to their legacy counterparts. Most users should keep Auto and only change this if they have specific memory constraints.

Warning

Changing the weight type requires re-loading the model, which can take a long time. Only change this setting if you understand the trade-offs.

---

📱 Platform Support

Platform	Status	Requirements
🤖 Android	✅ Supported	Android 11+ (API 30+) + with Vulkan 1.2
🪟 Windows	✅ Supported	Windows 10+ with Vulkan 1.2
🐧 Linux	✅ Supported	Vulkan 1.2+ drivers
🍎 macOS	✅ Supported	Metal support required
📱 iOS	✅ Supported	Metal support required

Tip

Memory Optimization:

• Android: Mmap is enabled by default. You can manually disable it in Settings if you encounter any issues.
• macOS: CLIP on CPU is enabled by default to prevent potential crashes during generation.

[!NOTE] Vulkan Performance: Vulkan is currently used as a general-purpose acceleration backend. While it ensures broad compatibility, generation speeds may not be fully optimized compared to native implementations.

---

🎨 Community Showcase

Created something amazing? We'd love to see it! Share your generation details (prompt, seed, model, etc.) to help others learn and create better art.

👉 Submit your creation here

---

🏗️ Architecture & Tech Stack

Core Technologies

graph TB
    A[Compose Multiplatform UI] --> B[Kotlin ViewModels]
    B --> C[Koin DI]
    C --> D[JNI Bridge]
    D --> E[C++ Native Layer]
    E --> F[stable-diffusion.cpp]
    F --> G[Vulkan/Metal Backend]

Loading

Technology Stack

• Language: Kotlin Multiplatform
• UI Framework: Compose Multiplatform
• Dependency Injection: Koin v4.1.1
• Navigation: Jetpack Navigation Compose
• Networking: Ktor 3.2.3 + Sandwich 2.1.2
• Image Loading: Coil3 v3.3.0
• Concurrency: Kotlin Coroutines
• Native Engine: stable-diffusion.cpp
• Native Engine ++: llama.cpp

---

🚀 Getting Started

Prerequisites

• Android: Android 11+ device with Vulkan 1.2 support
• Desktop: Windows/Linux/macOS with compatible graphics drivers
• Development: Android Studio Ladybug or later / IntelliJ IDEA

Installation

Option 1: Download Pre-built Release

1. Visit Releases
2. Download the appropriate package for your platform
3. Install and launch

Option 2: Build from Source

# Clone the repository
git clone https://github.com/Onion99/KMP-MineStableDiffusion.git
cd KMP-MineStableDiffusion

# Build for Desktop
./gradlew :composeApp:run

# Build for Android
./gradlew :composeApp:assembleDebug

First Run

1. Launch the app
2. Load your Stable Diffusion model (GGUF format)
3. Enter your text prompt
4. Click generate and watch the magic happen! ✨

---

📚 Documentation

• 📝 Changelog - Version history

---

🤝 Contributing

Contributions are welcome! Whether it's:

• 🐛 Bug reports
• 💡 Feature requests
• 📝 Documentation improvements
• 🔧 Code contributions

Please read our Contributing Guidelines before submitting PRs.

---

📄 License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

---

🙏 Acknowledgments

Special thanks to:

• leejet/stable-diffusion.cpp - Awesome C++ SD implementation
• ggerganov/llama.cpp - LLM inference framework
• JetBrains Compose Multiplatform - UI framework
• The entire Stable Diffusion community 💜

---

💙 Support This Project

If you find this project useful:

• ⭐ Star this repository
• 🐛 Report bugs and suggest features
• 🔀 Fork and contribute
• 📢 Share with others

---

📬 Contact

• Issues: GitHub Issues
• Discussions: GitHub Discussions

---