Local Dream is an Android application that runs Stable Diffusion image generation with Snapdragon NPU acceleration, while also supporting CPU and GPU inference as fallbacks.
The tool addresses the challenge of running computationally intensive generative AI models on mobile devices. It leverages Qualcomm's QNN SDK to execute models on Snapdragon NPUs, which are specialized processors designed for AI workloads. For devices without NPU support or as alternatives, the application can fall back to CPU or GPU inference using the MNN framework. The project supports Stable Diffusion 1.5 and SDXL models, with NPU acceleration available on Snapdragon processors with Hexagon V68 architecture or newer for SD1.5, and Snapdragon 8 Gen 3 and newer for SDXL. The application provides text-to-image, image-to-image, and inpainting capabilities.
Developers should choose this tool if they need to deploy Stable Diffusion on Android devices with hardware acceleration. It suits projects targeting Snapdragon-equipped Android phones where on-device inference is preferred over cloud-based generation. The project is particularly valuable for users with compatible Snapdragon NPUs, as NPU execution significantly reduces latency and power consumption compared to CPU or GPU inference. For devices without NPU support, CPU and GPU inference remain viable options, though SDXL models require substantial memory. The application is open source and free to use.
The project maintains active development with a community discussion channel for testing and support. Documentation is centralized on a dedicated guide site rather than in the repository itself. The codebase integrates multiple specialized libraries for different inference backends and includes image processing features like upscaling and NSFW content filtering built on established open-source foundations.
