Termux-Diffusion (Python)

Official Z-Image Turbo (6.0B DiT) & Pure Native On-Device Diffusion Acceleration for Android Termux & Samsung Galaxy.
Zero PRoot. Zero Virtualization. 100% Native ARM64 Bionic libc, Khronos Vulkan 1.1+ GPU Shaders & NEON SIMD Vector Engine.

🚀 Major Breakthrough: 6.0B DiT (Z-Image Turbo) Sovereign Mobile Execution

termux-diffusion v1.8.0 introduces official production support for Z-Image Turbo, empowering mobile devices to execute modern 6.0 Billion Parameter Diffusion Transformers (DiT) natively on-device.

🌟 Tri-Engine Asymmetric Architecture

• Text Encoder (CPU 4-Core): 4.0B LLM (Qwen3-4B-Instruct-2507-Q2_K.gguf) performs dense context embedding on CPU Big/Prime cores (clip_on_cpu=True).
• DiT Denoising Backbone (Vulkan GPU): 6.0B Diffusion Transformer (z_image_turbo-Q2_K.gguf) accelerated via Khronos Vulkan compute shaders with Tiled Flash Attention (diffusion_fa=True).
• Ultra-Fast VAE Decoder (CPU): 10MB Tiny AutoEncoder for FLUX.1 (taef1.safetensors) completes latent decoding rapidly without memory spikes (vae_on_cpu=True).
• Layer Streaming (stream_layers=True): Breaks through mobile memory limits by streaming layers into Vulkan VRAM sequentially, maintaining a strict 1.0 GB VRAM footprint (max_vram="vulkan0=1") and completely eliminating Android Low Memory Killer (LMK) aborts.

📸 Sovereign Physical Verification: Galaxy S21 DiT 8-Step Output

• Verification Image: Galaxy S21 Z-Image Turbo 8-Step PNG Artifact
• Verification Prompt (Ground Truth):

  "A cinematic photo of a neon cybernetic tiger walking in Seoul street at night"

• Verified Input Parameter Engineering:
  • steps=8 (Full numerical convergence ODE solver)
  • cfg_scale=1.0 (Optimal guidance for distilled DiT architecture)
  • sampler="euler" (1st-order Ordinary Differential Equation solver)
  • device="vulkan" (Physical GPU acceleration)
  • diffusion_fa=True (Tiled Flash Attention enabled)
  • stream_layers=True (Layer residency streaming over AXI bus)
  • max_vram="vulkan0=1" (Strict 1.0 GB VRAM ceiling, Zero-LMK protection)
  • clip_on_cpu=True (Offloads LLM prompt embedding to CPU Big/Prime cores)
  • vae_on_cpu=True (Prevents GPU VRAM allocation collisions during decoding)
  • taesd="auto" (Ultra-fast 10MB FLUX.1 latent decoder)
• Optical Realism: Specular light scattering from neon signboards on wet Seoul asphalt road surfaces, razor-sharp cybernetic armor plates, and micro-texture whiskers cleanly isolated from background noise.
• Runtime Stability: 100% completion achieved with zero Out-of-Memory (LMK) aborts and zero thermal throttling termination.

🔬 Academic Research & Lab Report Reference

• Full Technical Research Report: Galaxy S21 Z-Image Turbo Vulkan Research Report
• Academic Research Paper: AMEVA Labs | Sovereign On-Device AI Research, Newsletter & Discussion

💻 Python SDK Installation & Practical User Manual

1. Installation & Provisioning

pkg update && pkg install -y python clang termux-api
pip install termux-diffusion
termux-diffusion install

2. Python API Practical Code Example

import termux_diffusion as td

# 1. Official Z-Image Turbo 6.0B DiT Synthesis via Python API
image_path = td.generate(
    prompt="A cinematic photo of a neon cybernetic tiger walking in Seoul street at night",
    preset="z-image-turbo",
    steps=8,
    cfg_scale=1.0,
    sampler="euler",
    device="vulkan",
    diffusion_fa=True,
    stream_layers=True,
    max_vram="vulkan0=1",
    clip_on_cpu=True,
    vae_on_cpu=True,
    taesd="auto",
    output_path="/sdcard/Pictures/TermuxDiffusion/cyber_tiger_dit.png"
)
print(f"Generated via 6.0B DiT Tri-Engine: {image_path}")

# 2. Fast UNet Standard Inference (DreamShaper 8 LCM 6-Step)
lcm_path = td.generate(
    prompt="A cinematic photo of a neon cybernetic tiger walking in Seoul street at night",
    model="anime",
    steps=6,
    cfg_scale=1.5,
    sampler="lcm",
    device="vulkan",
    diffusion_fa=True,
    vae_on_cpu=True,
    output_path="/sdcard/Pictures/TermuxDiffusion/cyber_tiger_lcm.png"
)
print(f"Generated via DreamShaper 8 LCM: {lcm_path}")

🎛️ Python API Parameter Reference

| Argument | Type | Default | Engineering Purpose | | :--- | :---: | :---: | :--- | | prompt | str | Required | Text conditioning prompt | | width, height | int | 512 | Image dimensions (multiples of 64) | | steps | int | 8 (DiT) / 6 (LCM) | Numerical solver sampling iterations | | cfg_scale | float | 1.0 (DiT) / 1.5 (LCM) | Classifier-Free Guidance scale weight | | sampler | str | "euler" | Denoising ODE solver algorithm | | device | str | "vulkan" | Accelerator device ("vulkan", "cpu") | | preset | str | "z-image-turbo" | Verified configuration bundle | | stream_layers | bool | True (DiT) | Enables dynamic layer streaming over AXI bus | | max_vram | str | "vulkan0=1" | GPU memory ceiling (1.0 GB) to avert OS LMK aborts | | diffusion_fa | bool | True | Flash Attention $O(N)$ memory reduction | | clip_on_cpu | bool | True | Keeps LLM text encoder on CPU cores | | vae_on_cpu | bool | True | Offloads VAE decode from GPU to host CPU | | taesd | str | "auto" | Tiny AutoEncoder for ultra-fast latent decoding | | output_path | str | None | Path to save output PNG artifact |

🌐 Official Documentation & Ecosystem Links

• Official Architecture & API Reference
• Advanced Parameters Handbook
• GitHub Repository
• AMEVA Open-Source Foundation Portal

📄 License

Licensed under the Apache-2.0 License. Copyright (c) 2026 Eunho Kim (@uno-km).