Android native image upscaler powered by ncnn. Uses AI models to upscale images with high quality on-device, with Vulkan GPU acceleration.
Forked from Tumuyan/RealSR-NCNN-Android — rewritten in Kotlin with Jetpack Compose (Material 3), dynamic model downloading, result caching, and several stability improvements.
Screenshots taken on Mali-G52 | 4gb
<<<<<<< HEAD
- AI Image Upscaling — select from multiple ncnn models (Real-ESRGAN, RealSR, etc.)
- Vulkan GPU Acceleration — automatic Vulkan detection; falls back to CPU gracefully
- Progress Tracking — real-time ETA and percentage parsed from ncnn stdout
- Output Preview — pinch-to-zoom viewer with image metadata; save or share the result =======
- AI Image Upscaling — select from multiple ncnn models (Real-ESRGAN, etc.)
- Vulkan GPU Acceleration — automatic Vulkan detection; falls back gracefully
- Background Processing — foreground service keeps upscaling running while the app is in the background
- Batch Processing — upscale multiple images in one run
- Progress Tracking — real-time percentage, elapsed time and ETA from ncnn stdout
- Before/After Compare — slider to compare the original and upscaled image inline
- Output Preview — view and compare inline; save or share the result
f7dabff (i forgot, what this was about)
- Result Caching — in-memory LRU cache avoids redundant processing
- Noexec Bypass — runs binaries on noexec-mounted filesystems via
/proc/self/fd - Dynamic Asset Download — model
.bin/.paramfiles fetched from GitHub Releases on first launch - Debug Log — 500-entry ring buffer with copy-to-clipboard and file save
- Customizable — tile size, thread count, save format/name, notification mode, custom command arguments
- Material 3 UI — dark theme, bottom navigation, scrollable settings
- Android 8.0+ (API 26+)
- Vulkan-compatible GPU recommended (CPU fallback available)
Download the latest APK from the Releases page and install it.
Model assets are downloaded automatically on first launch.
Requires JDK 17 and Android SDK with NDK 27+ and CMake 3.22+.
./gradlew assembleRelease -Dorg.gradle.java.home=/path/to/jdk-17- Launch the app — model assets download on first run
- Tap the image card to select an image
- Choose a model from the dropdown
- Tap Run to upscale
- View the result in the output card; save or share from there
Tweak settings in the Settings tab: tile size, thread count, save format, naming scheme, custom arguments, and notification mode.
Place .bin + .param file pairs into the models/ folder. The app will detect and list them automatically.
Compatible sources:
- Real-ESRGAN releases — grab the
ncnnzip from assets - Upscayl model pack
- RealSR-NCNN-Android assets
To convert a PyTorch .pth model yourself: PyTorch → ONNX → ncnn using torch.onnx.export and onnx2ncnn.
- Black output on some Adreno GPUs —
vkQueueSubmit failed -4(VK_ERROR_DEVICE_LOST) is a known upstream ncnn issue on Qualcomm Adreno drivers. The GPU watchdog kills heavy shader dispatches before they complete. Workarounds: use a lighter model (x2 or anime variant), or reduce tile size in settings. Tracked upstream at Tencent/ncnn#925.
| Layer | Tech |
|---|---|
| Language | Kotlin + C++ |
| UI | Jetpack Compose / Material 3 |
| Inference | ncnn (Vulkan) |
| Image processing | ImageMagick |
| Image loading | Coil |
| Preferences | DataStore |
| Min SDK | API 26 (Android 8.0) |
MIT