AndroidCraft — on-device generative AI studio (chat, image, voice, video) + CI/CD APK

[Ishuin]

What this is

AndroidCraft — a fully on-device generative AI app for Android (Kotlin, Jetpack Compose, MVVM). Chat, generate images, synthesize speech, and make short video clips — all running locally on the phone, with models you download in-app. Every APK is built in CI, never on a device or laptop.

main is seeded at the initial scaffold commit; this PR contains the full build-out on top of it (media export, the image engine selector, real native + ONNX diffusion, and the curated model catalog).

Features

Tab	Runs on-device
Chat	Offline LLM inference via Google AI Edge / MediaPipe GenAI (LiteRT .task)
Image	Method selector: Procedural (instant) · Diffusion · SD.cpp (native stable-diffusion.cpp) · Diffusion · ONNX (ONNX Runtime, shown when a bundle is present)
Voice	Text-to-speech → WAV via the system neural TTS engine
Video	Prompt-seeded motion clip → real H.264 MP4 (MediaCodec/MediaMuxer)
Models	Curated one-tap downloads, sorted by device RAM fit, with a "Best for your device" badge

Every generated image, voice clip, and video can be saved (gallery / Music / Movies) and shared.

Curated models — zero manual setup

The Models screen offers only vetted, ungated, direct-download models — no account, no Hugging Face token, no links to find, no manual URL entry. They're matched to the device's reported RAM so the user doesn't have to guess what fits.

• Chat: Gemma 3 1B (int4) · Qwen 2.5 1.5B (int8) — the same LiteRT models Google's AI Edge Gallery ships.
• Image: Stable Diffusion 1.5 as single-file GGUF (Q8) for the native diffusion engine — one tap to download and run.

Real on-device diffusion

• stable-diffusion.cpp is vendored as a submodule under app/src/main/cpp and compiled for arm64 via NDK/CMake; sdjni.cpp is a thin JNI bridge. The whole pipeline (tokenizer, scheduler, UNet, VAE) runs in C++.
• ONNX Runtime engine implements the full SD 1.5 pipeline in Kotlin (CLIP byte-level BPE tokenizer + UNet with classifier-free guidance + DDIM scheduler + VAE) on onnxruntime-android.

CI/CD

• .github/workflows/android-build.yml — on every build-affecting commit: inits native submodules, sets up JDK 17 + Android SDK + NDK 27 / CMake, runs unit tests, compiles the native library, assembles the debug APK, and uploads it as the androidcraft-debug-apk artifact.
• .github/workflows/release.yml — on a v* tag, builds and publishes a GitHub Release with the APK (auto-signs if keystore secrets are set).

Notes / honest caveats

• The app builds and links cleanly in CI. The diffusion engines are compile-verified in CI but not device-tested from the build environment — the first real generation on-device is the true test. Each engine is gated behind "is the model present," so a mismatch shows an error on that one tab rather than crashing the app.
• Toolchain: AGP 8.7.3 / Kotlin 2.0.21 / compileSdk 35 / minSdk 26 / arm64.
• Clone with --recurse-submodules for local builds.

🤖 Generated with Claude Code

---

Generated by Claude Code