Termux-Diffusion

Official Z-Image Turbo (6.0B DiT) & Pure Native On-Device Diffusion Acceleration for Android Termux

1-Line Quick Installation

Install the official package directly into your runtime:

pip install termux-diffusion && termux-diffusion install
# or:
npm install termux-diffusion && npx termux-diffusion install

The Engineering Challenge

Executing modern 6.0B Diffusion Transformers on mobile hardware requires cloud dependencies or leads to immediate Out-of-Memory (LMK) aborts due to VRAM exhaustion.

The Architectural Breakthrough

Executes 6.0B DiT (Z-Image Turbo) and quantized Stable Diffusion models with Tri-Engine asymmetric offloading, layer streaming, Vulkan compute shaders, and ARMv8.2-A DotProd/FP16 NEON SIMD.

Key Capabilities & Built-in Hardening

Official Z-Image Turbo (6.0B DiT)

Native mobile execution of 6.0B Diffusion Transformer via Tri-Engine pipeline (Qwen3-4B LLM text encoder + DiT backbone + 10MB fast VAE).

Dynamic Layer Streaming & Zero-OOM

Streams DiT layers into Vulkan VRAM sequentially, maintaining a strict 1.0 GB VRAM footprint to prevent mobile LMK kills.

Dual-Engine Vulkan GPU & CPU Baseline

Bypasses PRoot with precompiled Khronos Vulkan 1.1+ compute kernels and OpenMP NEON SIMD vector runtime.

Android MediaStore Auto-Indexing

Automatically synchronizes generated images into Pictures/TermuxDiffusion and registers with Android MediaStore.

Supported Compute Kernels & Operations

Subsystem Category	Operations & Kernels	Status
Diffusion Transformer (DiT)	Z-Image Turbo (6.0B DiT), Qwen3-4B-Instruct LLM Text Encoder, TAESD FLUX.1 VAE	Production
Diffusion UNet Engines	SD 1.5, SDXS Distilled, SD Turbo, DreamShaper 8 (LCM 6-step), Realistic Vision, GGML Quantized Weights	Production
Compute Acceleration	Khronos Vulkan 1.1+ GPU Shaders, Dynamic Layer Streaming, ARM64 DotProd/FP16 NEON SIMD, OpenMP	Production
Media Integration	PNG/WebP Export, Android MediaStore Intent Broadcast	Production

Canonical Usage Example

import termux_diffusion as td

# Generate image using official on-device acceleration
image_path = td.generate(
    prompt="A cinematic photo of a neon cybernetic tiger walking in Seoul street at night",
    model="anime",
    steps=6,
    device="vulkan"
)
print(f"Generated and saved to Gallery: {image_path}")

import { generate } from "termux-diffusion";

const result = await generate({
  prompt: "A cinematic photo of a neon cybernetic tiger walking in Seoul street at night",
  model: "anime",
  steps: 6,
  device: "vulkan"
});
console.log("Image created:", result.outputPath);