Ship Stable Diffusion in an Android App

What you'll learn

• ✓Implement a Stable Diffusion sampling loop, scheduler and tokenizer in Kotlin with ONNX Runtime Mobile
• ✓Build the same feature in React Native with a native pixel encoder
• ✓Select execution providers per platform and keep the UI thread responsive
• ✓Build an offline assistant that chains an LLM prompt expansion into image generation
• ✓Benchmark native versus React Native fairly and choose a track with data

About this course

On a phone there is no convenient pipeline object: you write the sampling loop, wire the tokenizer, manage tensor lifetimes and stream progress to the user yourself. This course builds that image app twice — in Kotlin and in React Native — and finishes with a capstone and a benchmark.

Track A is Kotlin with ONNX Runtime Mobile: project and Gradle setup, an SdEngine that runs text encoder, U-Net and VAE decoder, the Euler scheduler implemented from the equations, a CLIP byte-pair tokenizer, a ViewModel with streaming progress, a Compose screen, and the NNAPI versus QNN decision. Track B is React Native with onnxruntime-react-native, including a native pixel encoder on Android and iOS and the pain points of the cross-platform path. Then cross-platform acceleration: per-platform execution-provider selection, avoiding JS-thread freezes, iOS Instruments and Perfetto. The capstone chains a small language model that expands the prompt into the image pipeline, with a safety filter and checkpoint resume. The final session is a dual-track benchmark: metrics that are fair, in-app hooks, a CI performance gate, results analysis, and a decision framework for choosing a track.

The code is read by intent, and the narration lists explicitly what must be fixed in the sample code before shipping.

**Licensing — read before you build a business on this.** Putting model weights inside an app is redistribution, so this course ships a `Resources/LICENSES.md` with every asset checked against its actual licence text. Stable Diffusion 1.5 is CreativeML Open RAIL-M: commercial use is allowed, but you must carry its use restrictions through to your own users as an enforceable term. The capstone's image model, **SDXL Turbo, is under the Stability AI Community License — free for commercial use only while you are under US $1,000,000 annual revenue**, and it requires a "Powered by Stability AI" attribution. Gemma 2, which expands the capstone's prompts, requires you to hand every user a copy of the Gemma Terms and to carry its Prohibited Use Policy into your own terms.

Curriculum6 sections · 14 lectures · 1h 55m

Read alongside the course