Modelpublic

alphastack1/bonsai-image

sourceHugging Faceotherupdated 4mo agoView on Hugging Face

1likes

Model Card

Bonsai Image

On-device AI image generation for Android. A single APK that runs a diffusion model entirely on your phone's GPU via WebGPU — no accounts, no servers, no cloud.

This repo hosts the all-inclusive APK (~2.5 GB, model bundled inside) because it's over GitHub's 2 GB release-asset limit.

Download

[⬇️ Bonsai-Image.apk](https://huggingface.co/alphastack1/bonsai-image/resolve/main/Bonsai-Image.apk) (~2.5 GB)

1.Enable Install from unknown sources on your Android device.
2.Download and open the APK.
3.Type a prompt (or attach a photo for image-to-image) and tap Generate.

Requirements

—Android 12+ with a recent Android System WebView / Chrome (121+)
—A Qualcomm Adreno or ARM Mali GPU (most modern phones)
—~3 GB free storage
—WebGPU support in the system WebView — the app reports this in Settings → Device

How it works

Layer	Detail
UI	Single-file WebView app, dark theme, on-device history (localStorage)
Engine	SD-Turbo via `onnxruntime-web` on the WebGPU execution provider
Model	fp16 ONNX (text encoder + UNet + VAE), bundled in the APK
Tokenizer	CLIP, via `@xenova/transformers`

Generation runs locally. First launch fetches a small (~25 MB) GPU runtime once, then it's fully on-device.

Status — experimental

This is an early build. The current engine is SD-Turbo, used as a working stand-in. PrismML's Bonsai 1-bit/ternary image model (a 1-bit quantization of FLUX.2 Klein) has no Android/WebGPU runtime yet; when one ships, it drops into the same engine slot and the app around it stays identical.

Installs alongside the separate Bonsai LLM text app — different package, violet icon.

License

The bundled SD-Turbo weights are governed by the Stability AI Non-Commercial Research Community License (see SD-Turbo). Non-commercial use only.