File size: 3,610 Bytes
80344a7 a3d9e9b 80344a7 a3d9e9b 80344a7 a3d9e9b 80344a7 a3d9e9b 80344a7 a3d9e9b 80344a7 a3d9e9b 80344a7 a3d9e9b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 | # Staticplay SD1 Mobile (On-Device, Android)
Goal: ship a **real offline** Android app (no servers, no cloud) that runs an SD1-class model **on-device** using **NNAPI** (via a native runtime).
This repo currently includes a working offline SD1.5 ONNX "pack" runner (unzip + inspect + generate).
## Key points (no surprises)
- Use a **prebuild** / **native build** (Expo Go won't cut it for native inference).
- "Offline" means the **model files are on the phone** (bundled in the APK or copied into app storage).
- ONNX Runtime is integrated via a tiny **Android native module** (no JSI auto-install).
## Run (Android)
```bash
cd /home/tiny/projects/staticplay-sd1-mobile
export ANDROID_HOME=/home/tiny/Android/sdk
export PATH="$ANDROID_HOME/platform-tools:$PATH"
npx expo prebuild --clean
npx expo run:android # debug (uses Metro)
```
For a fully offline build (no Metro), install the release build:
```bash
cd /home/tiny/projects/staticplay-sd1-mobile/android
NODE_ENV=production ./gradlew :app:installRelease
```
Then press "Test native ONNX (NNAPI -> CPU)" in the app.
## Bundle With Expo (Windows, release APK)
Use these commands from this project root.
```powershell
# 1) Install dependencies (once)
npm install
# 2) Regenerate native Android project if config/plugins changed
npx expo prebuild --platform android --clean
# 3) Build release APK
cd android
./gradlew.bat :app:assembleRelease
# 4) Install to connected phone (USB debugging on)
./gradlew.bat :app:installRelease
```
APK output path:
`android/app/build/outputs/apk/release/app-release.apk`
Notes:
- This v2 build uses package id `com.anonymous.staticplaysd1mobile.v2`, so v1 can stay installed.
- Some setup buttons can take longer than expected (especially inspect actions).
- Generation works offline but is still being optimized for speed.
- Website: `https://staticplay.co.uk`
## SD1.5 pack + generation (offline)
1) Put the SD1.5 ONNX pack zip at the app external directory as:
`/storage/emulated/0/Android/data/com.anonymous.staticplaysd1mobile/files/sd15-onnx-pack.zip`
2) In the app:
- (Optional) `Show app external pack path` (for sanity check)
- `Unzip SD15 pack from external` (one-time; can take minutes)
- `Load tokenizer (bundled)` (no external files required)
- Toggle backend:
- `Backend: NNAPI (experimental)` = fastest, can be unstable on some devices/drivers
- `Backend: CPU (safe)` = slower, but avoids NNAPI driver issues
- `Generate 512x512 (quick 4)` to prove the end-to-end pipeline
- `Generate 512x512 (quality 20)` for higher quality (slower)
3) Output images save to the same app external folder as `sd_out_*.png`.
## SD1 plan (high level)
1) Pick an SD1 distilled variant (LCM/Turbo-style) to keep steps ~4-8.
2) Export components to ONNX (text encoder, UNet, VAE decoder).
3) Quantize mostly UNet (INT8) + keep VAE in FP16/BF16.
4) Run inference via a native runtime with NNAPI; fall back to CPU if NNAPI can't compile a node.
5) Ship the ONNX files in-app (or as offline downloadable packs).
## V2 notes (2026-02-12)
- This branch installs as a separate app id (`com.anonymous.staticplaysd1mobile.v2`) so v1 stays on device.
- Guided mode is enabled to show one button at a time for easier setup flow.
- Some actions can take longer than expected, especially:
- `Inspect pack UNet`
- `Inspect pack text_encoder`
- `Inspect pack vae_decoder`
- first unzip/generation after launch
- Overall generation is currently functional but still slower than target. Ongoing work is focused on speed improvements.
- Public site: `https://staticplay.co.uk`
|