duration-ai/bonsai-image-android

This repo demonstrates running a compact AI image-generation model entirely on a 2025 Android flagship phone, with no internet connection required. The model produces 512x512 images from text prompts, and all processing happens on the device's specialized chips. The model is called Bonsai Image, built by PrismML from a 4-billion-parameter architecture called FLUX.2 klein, compressed down using a technique called ternary quantization (weights stored as three possible values instead of full floating-point numbers). The image generation happens in three stages: the text prompt is encoded into a numerical representation on the phone's main processor, a diffusion transformer then refines a noisy image over four steps on the phone's dedicated AI chip (Qualcomm's Hexagon NPU), and finally the result is decoded back into a visible image on the CPU again. The Hexagon NPU path is what makes this interesting. Running the transformer on the CPU would take roughly eight to nine minutes per image, which is not practical. Running it on the NPU brings that down to about two minutes and twenty seconds for the full four-step render. The GPU was tested but crashed or faulted at the 512x512 size. Benchmark numbers for all three compute paths are documented in detail. Building this yourself requires the Qualcomm QAIRT SDK, an Android NDK, and a device with a Hexagon V79 chip (a 2025 Snapdragon 8 Elite phone). The process involves exporting the model's 27 individual blocks into compiled binaries for the NPU, cross-compiling a small C++ runner that chains them together on-device, and patching a CPU-side tool to hand off the text encoding and image decoding stages correctly. The total bundle size is about 10.7 GB, most of which is the compiled NPU binaries. This is a companion to the same team's iOS version, which runs the same model on a 2020 iPhone GPU through Apple's MLX framework. Both achieve roughly the same total render time. The repo serves as a technical reference for running large AI models on mobile hardware without cloud infrastructure.