local-dream Run Stable Diffusion on Android De @codeKK AndroidOpen Source Website

local-dream

Introduction: Run Stable Diffusion on Android Devices with Snapdragon NPU acceleration. Also supports CPU/GPU inference.

More: Author ReportBugs OfficialWebsite

Tags:

Android Stable Diffusion with Snapdragon NPU acceleration
Also supports CPU/GPU inference

About this Repo

This project is now open sourced and completely free. Hope you enjoy it!

If you like it, please consider sponsor this project.

[!NOTE] Currently focus on SD1.5 and SDXL models. SD2.1 models are no longer maintained due to poor quality and not popular. SDXL models are only available for Snapdragon 8 Gen 3, Snapdragon 8 Elite and Snapdragon 8 Elite Gen 5/ 8 Gen 5.

Most users don't get how to properly use highres mode for SD1.5. Please check here.

Now you can import your own NPU models converted using our easy-to-follow NPU Model Conversion Guide. And you can also download some pre-converted models from xororz/sd-qnn or xororz/sdxl-qnn or Mr-J-369. For SD1.5 models, download _min if you are using non-flagship chips. Download _8gen1 if you are using 8gen1. Download _8gen2 if you are using 8gen2/3/4/5. We recommend checking the instructions on the original model page to set up prompts and parameters.

You can join our telegram group for discussion or help with testing.

🚀 Quick Start

Download: Get the APK from Releases or Google Play(NSFW filtered)
Install: Install the APK on your Android device
Select Models: Open the app and download the model(s) you want to use

✨ Features

🎨 txt2img - Generate images from text descriptions
🖼️ img2img - Transform existing images
🎭 inpaint - Redraw selected areas of images
custom models - Import your own SD1.5 models for CPU (in app), SD1.5 and SDXL for NPU (follow conversion guide). You can get some pre-converted models from xororz/sd-qnn or xororz/sdxl-qnn or Mr-J-369
lora support - Support adding LoRA weights to custom CPU models when importing.
prompt weights - Emphasize certain words in prompts. E.g., (masterpiece:1.5). Same format as Automatic1111
embeddings - Support for custom embeddings like EasyNegative. SafeTensor format is required. Convert pt to safetensors using this
upscalers - 4x upscaling with realesrgan_x4plus_anime_6b and 4x-UltraSharpV2_Lite

🔧 Build Instructions

Note: Building on Linux/WSL is recommended. Other platforms are not verified.

Prerequisites

The following tools are required for building:

Rust - Install rustup, then run:

# rustup default stable
rustup default 1.84.0 # Please use 1.84.0 for compatibility. Newer versions may cause build failures.
rustup target add aarch64-linux-android

Ninja - Build system
CMake - Build configuration

1. Clone Repository

git clone --recursive https://github.com/xororz/local-dream.git

2. Prepare SDKs

Download QNN SDK: Get QNN_SDK_2.39 and extract
Download Android NDK: Get Android NDK and extract
Configure paths:
- Update QNN_SDK_ROOT in app/src/main/cpp/CMakeLists.txt
- Update ANDROID_NDK_ROOT in app/src/main/cpp/CMakePresets.json

3. Build Libraries

Linux

cd app/src/main/cpp/
bash ./build.sh

Windows

# Install dependencies if needed:
# winget install Kitware.CMake
# winget install Ninja-build.Ninja
# winget install Rustlang.Rustup

cd app\src\main\cpp\

# Convert patch file (install dos2unix if needed: winget install -e --id waterlan.dos2unix)
dos2unix SampleApp.patch
.\build.bat

macOS

# Install dependencies with Homebrew:
# brew install cmake rust ninja

# Fix CMake version compatibility
sed -i '' '2s/$/ -DCMAKE_POLICY_VERSION_MINIMUM=3.5/' build.sh
bash ./build.sh

4. Build APK

Open this project in Android Studio and navigate to: Build → Generate App Bundles or APKs → Generate APKs

Technical Implementation

NPU Acceleration

SDK: Qualcomm QNN SDK leveraging Hexagon NPU
Quantization: W8A16 static quantization for optimal performance
Resolution: For SD1.5, fixed 512×512 base model shape, more shapes available using patches. For SDXL, fixed 1024x1024 model shape
Performance: Extremely fast inference speed

CPU/GPU Inference

Framework: Powered by MNN framework
Quantization: W8 dynamic quantization
Resolution: Flexible sizes (128×128, 256×256, 384×384, 512×512)
Performance: Moderate speed with high compatibility

NPU High Resolution Support

[!IMPORTANT] Please note that quantized high-resolution(>768x768) SD1.5 models may produce images with poor layout. We recommend first generating at 512 resolution (optionally you can upscale it), then using the high-resolution model for img2img (which is essentially Highres.fix). The suggested img2img denoise_strength is around 0.8. After that, you can get images with better layout and details.

Device Compatibility

NPU Acceleration Support

SD1.5 compatible with devices featuring:

Snapdragon 8 Gen 1/8+ Gen 1
Snapdragon 8 Gen 2
Snapdragon 8 Gen 3
Snapdragon 8 Elite
Snapdragon 8 Elite Gen 5/8 Gen 5
Non-flagship chips with Hexagon V68 or above (e.g., Snapdragon 7 Gen 1, 8s Gen 3)

SDXL compatible with devices featuring:

Snapdragon 8 Gen 3
Snapdragon 8 Elite
Snapdragon 8 Elite Gen 5/8 Gen 5

Note: Other devices cannot download NPU models

CPU/GPU Support

RAM Requirement: ~2GB available memory
Compatibility: Most Android devices from recent years

Available Models

The following models are built-in and can be downloaded directly in the app:

Model	Type	CPU/GPU	NPU	Clip Skip	Source
SDXL Base 1.0	SDXL	❌	✅	-	HuggingFace
Illustrious v16	SDXL	❌	✅	-	CivitAI
AnythingV5	SD1.5	✅	✅	2	CivitAI
ChilloutMix	SD1.5	✅	✅	1	CivitAI
Absolute Reality	SD1.5	✅	✅	2	CivitAI
QteaMix	SD1.5	✅	✅	2	CivitAI
CuteYukiMix	SD1.5	✅	✅	2	CivitAI

🎲 Seed Settings

Custom seed support for reproducible image generation:

CPU Mode: Seeds guarantee identical results across different devices with same parameters
GPU Mode: Results may differ from CPU mode and can vary between different devices
NPU Mode: Seeds ensure consistent results only on devices with identical chipsets