Stable Diffusion on Android that finally uses the NPU

What it does

Local Dream is an open-source Android app for generating and editing images with Stable Diffusion entirely on-device. It uses Qualcomm’s QNN SDK to run inference on the Snapdragon Hexagon NPU where available, falling back to CPU or GPU execution via Alibaba’s MNN framework for unsupported chips. The app supports txt2img, img2img, and inpainting, and bundles Real-ESRGAN and UltraSharpV2 upscalers plus an opennsfw2-based content filter.

The interesting bit

Most mobile diffusion tools either phone home to a cloud API or hammer the GPU; this one actually targets the NPU. SD1.5 runs on Hexagon V68 and newer, while SDXL demands a Snapdragon 8 Gen 3 or above—older hardware is stuck with CPU/GPU for SD1.5 only. The authors also dropped SD2.1 entirely, bluntly citing poor quality and limited popularity, which saves the codebase from zombie-model maintenance.

Key highlights

• NPU acceleration for SD1.5 (Hexagon V68+) and SDXL (Snapdragon 8 Gen 3+)
• CPU/GPU fallback via MNN for chips outside the NPU support window
• On-device upscaling and built-in NSFW filtering
• Jetpack Compose UI with Material Design
• Fully free and open source, though documentation now lives on an external guide site

Caveats

• SDXL NPU support is restricted to very recent flagship Snapdragon chips
• SD2.1 support has been permanently removed
• In-repo documentation is minimal; guides are hosted externally

Verdict

Worth a look if you have a recent Android device and want local, offline image generation without cloud dependencies. Temper expectations if your phone is older than the Hexagon V68 era or if you need SDXL on budget hardware.