On-device object & watermark eraser for Android — tap it, and it's gone.
Vanish grew out of WatermarkRemover-AI-Revamped and is now its own app. Point it at any photo, tap the thing you want gone, and an AI model rebuilds the background behind it — all on the phone, no upload.
Obtainium (recommended) — tracks this GitHub repo's releases and auto-updates the app for you, no store account needed:
- Install Obtainium if you don't have it.
- Tap the badge above, or in Obtainium choose Add App and paste
https://github.com/SpyrosTheBoss/Vanish. - Obtainium picks up the latest signed release automatically, and again on every future tag.
Manual APK — grab the signed APK straight from the latest release and sideload it. You'll need to update it by hand when a new version ships.
Vanish isn't on Google Play or F-Droid — the bundled ML models push it over F-Droid/IzzyOnDroid's size limits, and their inclusion policy is opposed to AI-assisted development, which this project doesn't hide from.
- Tap-to-select — one tap and MobileSAM outlines the exact object. Lasso and brush are there too when you want manual control.
- On-device inpainting — MI-GAN (bundled, ~2s) rebuilds the background. No photo ever leaves the device.
- Choose your model — switch to LaMa in Settings for higher quality; it downloads on demand (208 MB).
- Before / after — drag to compare, then save to your gallery or share.
- Material 3 — dynamic color (Material You), light/dark, themed icon.
| Stage | Model | Runs |
|---|---|---|
| Select (tap) | MobileSAM (TinyViT encoder + decoder, ONNX) | once per photo + per tap |
| Erase | MI-GAN or LaMa (ONNX Runtime Mobile) | per removal |
Both stages run locally through ONNX Runtime. The scripts/ folder documents how each model was exported and verified.
Open the Vanish/ folder in Android Studio (JDK 17+, minSdk 29) and run. The MI-GAN and MobileSAM models are bundled in app/src/main/assets/; LaMa is fetched at runtime the first time you select it.
Inpainting models via IOPaint (MI-GAN, LaMa). Segmentation via MobileSAM. LaMa ONNX from Carve/LaMa-ONNX.
GPL-3.0. Model credits: MI-GAN and LaMa via IOPaint, segmentation via MobileSAM.