xororz/local-dream

What it solves

Local Dream provides a way to run Stable Diffusion image generation models locally on Android devices, eliminating the need for cloud-based services. It specifically optimizes performance by leveraging the Snapdragon NPU (Neural Processing Unit) for faster, more efficient inference on compatible hardware.

How it works

The application uses a combination of C++ libraries and Android-native UI frameworks. It utilizes the Qualcomm QNN SDK for NPU execution, Alibaba MNN for CPU execution, and stable-diffusion.cpp for performance optimizations. It supports SD1.5 and SDXL models, with NPU acceleration available for specific Snapdragon architectures (Hexagon V68+ for SD1.5 and Snapdragon 8 Gen 3+ for SDXL).

Who it’s for

Android users who want to generate AI images locally on their mobile devices without relying on an internet connection or external servers.

Highlights

• Hardware Acceleration: Leverages Snapdragon NPUs for high-performance local inference.
• Model Support: Compatible with Stable Diffusion 1.5 and SDXL.
• Model Flexibility: Supports CPU and GPU inference as a fallback or alternative to NPU acceleration.
• Integrated Tools: Includes image upscaling via Real-ESRGAN and UltraSharpV2, and NSFW content filtering via opennsfw2.