Termux-Diffusion

Native On-Device Stable Diffusion Runtime for Android Termux & Samsung Galaxy via Direct Bionic libc & Vulkan Compute Acceleration.
Zero PRoot. Zero Virtualization. 100% Native ARMv8.2-A NEON SIMD & Hardware GPU Acceleration.

📑 Table of Contents

1. Overview & Key Capabilities
2. Installation Guide
3. Enabling Hardware GPU Acceleration (with ameva-runtime)
4. Basic Usage (CLI, Python, Node.js)
5. Advanced Workflows
6. Feature & Parameter Matrix
7. Production Code Examples & Self-Diagnostics
8. Real-World Outputs & Hardware Benchmarks
9. GPU Interconnect Architecture & Compatibility Matrix (Adreno vs. Mali)
10. Hardware Requirements & Operational Limits
11. 24/7 Unattended Background Execution Guide (Termux -> Android -> ADB)
12. License & Permissible Use
13. Keywords & Discoverability Index

1. Overview & Key Capabilities

termux-diffusion is a production-grade, on-device diffusion inference engine engineered specifically for Android smartphones. Unlike legacy approaches relying on slow PRoot or chroot virtual machines, termux-diffusion compiles directly against Android's native Bionic libc and binds directly to host GPU drivers to execute high-quality 512x512 image synthesis natively on edge devices.

• Native Bionic libc ABI Direct Binding: Runs directly inside Termux with zero virtual memory indirection, achieving bare-metal compute efficiency.
• Dual Compute Acceleration: Integrates ARMv8.2-A DotProd/FP16 vector instructions with Qualcomm Adreno and ARM Mali Vulkan compute pipelines.
• Built-in VAE Tiling: Eliminates the 1.2 GB memory spike during latent-to-pixel decoding, reducing peak RAM consumption by ~70% and preventing kernel Out-Of-Memory (OOM) aborts.
• Automated Android MediaStore Indexing: Synchronizes generated images directly into Pictures/TermuxDiffusion and the native Samsung Gallery app in real time.
• Autonomous WakeLock Lifecycle Protection: Automatically acquires an Android kernel CPU WakeLock during active inference to prevent thermal clock throttling when the display turns off.

2. Installation Guide

2.1 Termux System Prerequisites

Launch the Termux terminal and install required native compilers, image processing libraries, and Vulkan tools:

pkg update && pkg install -y python nodejs clang git libjpeg-turbo libpng termux-api vulkan-tools

2.2 Package Installation (Python & Node.js)

• Python SDK (PyPI):

  pip install termux-diffusion
  

• Node.js SDK & CLI (NPM):

  npm install -g termux-diffusion
  

2.3 One-Command Native Engine Provisioning

Run the automated installer to detect your device architecture and provision prebuilt native binaries or compile on-device:

termux-diffusion install

3. Enabling Hardware GPU Acceleration (with ameva-runtime)

To unlock mobile GPU acceleration via Vulkan compute shaders and achieve significant speedups over pure CPU execution, install termux-diffusion alongside ameva-runtime in a single command:

🌟 One-Line Installation

# Python SDK
pip install termux-diffusion ameva-runtime

# Node.js SDK
npm install -g termux-diffusion @ameva/runtime

🔮 Acceleration Mechanics with ameva-runtime

1. Dynamic Driver Probing: Automatically detects the underlying SoC (Snapdragon vs. Exynos/Dimensity) and dynamically locates the vendor Bionic Vulkan ICD (/system/lib64/libvulkan.so) in < 1ms.
2. Vendor-Tailored Pipeline Dispatch: Automatically binds optimized SPIR-V compute shaders tailored for Qualcomm Adreno or ARM Mali architectures without manual driver compilation.
3. big.LITTLE Core Affinity Governor: Binds worker compute threads exclusively to high-performance prime cores (Cortex-X / Cortex-A78) while streaming command buffers to the GPU queue.

Verify GPU driver detection and hardware readiness:

termux-diffusion doctor

4. Basic Usage (CLI, Python, Node.js)

4.1 Terminal CLI

# Standard Photorealistic Generation (DreamShaper v8 Q4_0 preset)
termux-diffusion generate "Cyberpunk Seoul street at night, neon lights, 8k, photorealistic"

# Ultra-Fast Generation (SDXS-512-0.9 1~4 steps convergence) with GPU
termux-diffusion generate "Cute fluffy white cat with sapphire eyes on the beach" -m speed --gpu

# Explicit Output File Specification
termux-diffusion generate "A majestic snow tiger in winter forest" -o /sdcard/tiger.png

4.2 Python SDK

import termux_diffusion as td

# Generate high-fidelity image on mobile hardware
result = td.generate(
    prompt="Cinematic portrait of an astronaut floating in colorful nebula, 8k, masterpiece",
    negative_prompt="lowres, bad anatomy, deformed, blurry, artifacts",
    model="realistic",       # 'realistic' | 'speed' | 'turbo' | 'anime'
    device="gpu",            # 'gpu' | 'cpu' | 'auto'
    steps=10,                # Denoising iterations
    cfg_scale=4.5,           # Classifier-Free Guidance scale
    width=512,
    height=512,
    seed=-1                  # -1 for random seed
)

print(f"Generated Image: {result.path}")
print(f"Samsung Gallery Path: {result.gallery_path}")
print(f"Inference Time: {result.elapsed_sec:.1f}s")

4.3 Node.js / TypeScript SDK

import { generate } from "termux-diffusion";

async function main() {
  const result = await generate({
    prompt: "An ancient temple hidden in a lush rainforest with golden sunlight beams, 8k",
    negativePrompt: "blurry, low quality, dark, distorted",
    model: "realistic",
    device: "gpu",
    steps: 10,
    cfgScale: 4.5,
    width: 512,
    height: 512
  });

  console.log(`Success: ${result.path} (${result.elapsedSec}s elapsed)`);
}

main();

5. Advanced Workflows

5.1 Image-to-Image (Img2Img Transformation)

Transform an existing photograph or sketch into stylized AI artwork:

result = td.generate(
    prompt="Futuristic robotic mecha warrior with glowing blue armor, 8k",
    init_img="source_sketch.png",
    strength=0.65,            # Denoising strength (0.0 keeps source, 1.0 full reimagining)
    model="realistic"
)

# Terminal CLI
termux-diffusion generate "Robotic mecha warrior" -i source_sketch.png --strength 0.65

5.2 VAE Tiling (Mobile OOM Prevention)

Standard VAE decoding creates an instantaneous 1.2 GB RAM spike when converting latents to RGB pixels at 768x768 or higher resolutions. Enabling vae_tiling processes the latent tensor in spatial chunks, lowering peak memory consumption by 70%:

result = td.generate(
    prompt="Breathtaking wide landscape of Alpine mountains during sunset",
    width=768,
    height=768,
    vae_tiling=True           # Enforces low-memory tiled VAE decode
)

5.3 TAESD (Tiny AutoEncoder) Ultra-Fast Decoding

Replaces heavy multi-layer autoencoders with Tiny AutoEncoder for Stable Diffusion, slashing the final decode step from 12 seconds down to < 0.1 seconds:

result = td.generate(
    prompt="A cute golden retriever puppy sitting in a flower basket",
    model="speed",
    taesd="taesd.gguf",       # Path to TAESD weights
    steps=4
)

5.4 LoRA Style Adaptation & ControlNet Guidance

result = td.generate(
    prompt="A cyberpunk warrior swinging an energy katana, <lora:cyber_armor:0.8>",
    lora_dir="/data/data/com.termux/files/home/loras",
    control_net="controlnet-canny.gguf",
    control_image="edge_guide.png",
    control_strength=0.9
)

5.5 High-Fidelity Sampler & Scheduler Pairings

# Optimal pairing for hyperrealistic skin textures and micro-details: DPM++ 2M + Karras
result = td.generate(
    prompt="Studio portrait of an elderly watchmaker working with intricate gears",
    sampling_method="dpm++2m",
    schedule="karras",
    steps=14,
    cfg_scale=5.0
)

6. Feature & Parameter Matrix

7. Production Code Examples & Self-Diagnostics

7.1 Automated Continuous Batch Loop

import termux_diffusion as td

prompts = [
    "A cozy rainy cafe street in Kyoto, watercolor style",
    "A futuristic cyberpunk police car chasing a neon drone",
    "A serene zen garden with blooming cherry blossoms, golden hour"
]

for idx, p in enumerate(prompts):
    print(f"[{idx+1}/{len(prompts)}] Generating: {p}")
    res = td.generate(prompt=p, model="speed", steps=4, width=512, height=512)
    print(f" -> Saved to: {res.path} ({res.elapsed_sec:.1f}s)")

7.2 Asynchronous Non-Blocking Execution (FastAPI / Bot Integration)

import asyncio
from termux_diffusion import generate_async

async def main():
    print("Dispatching asynchronous diffusion task...")
    task = asyncio.create_task(
        generate_async(
            prompt="A majestic eagle soaring above snow-capped Rocky Mountains",
            model="realistic",
            steps=10
        )
    )
    # Concurrent I/O operations continue unblocked
    await asyncio.sleep(1)
    print("Event loop running freely without thread lock...")
    
    result = await task
    print(f"Task completed: {result.path}")

asyncio.run(main())

7.3 Hardware Diagnostic Inspection

import termux_diffusion as td

profile = td.get_hardware_profile()
print(f"CPU Architecture: {profile.cpu_arch}")
print(f"GPU Device: {profile.gpu_name}")
print(f"Vulkan Hardware Available: {profile.vulkan_available}")
print(f"Optimal Backend: {profile.recommended_backend}")

8. Real-World Outputs & Hardware Benchmarks

All benchmark metrics and rendered outputs represent physical executions on commercial Samsung Galaxy smartphones.

🖼️ Real-Device Rendered Samples

DreamShaper v8 (10 Steps)	SDXS-512 (1 Step Fast)	SD 1.5 Native (4 Steps)
Galaxy S25 + Vulkan (32s)	Galaxy S21 + SDXS (7.2s)	Galaxy S20 + Turbo (18s)

📊 Physical Benchmark Matrix (512x512 Resolution)

9. GPU Interconnect Architecture & Compatibility Matrix

9.1 Native Bionic libc Binding Mechanism

Virtual machine abstractions (e.g. PRoot) incur severe context switching and memory copy penalties when communicating with Linux kernel GPU drivers. termux-diffusion bypasses user-space shims and loads the host Android Bionic ICD directly:

[termux-diffusion Native Core]
              │
              ▼
       (Direct dlopen)
              │
              ├──> /system/lib64/libvulkan.so   (Host Android System ICD - Primary)
              └──> /vendor/lib64/libvulkan.so   (Vendor Hardware Driver)

Critical Safety Rule: Dynamically linking Termux's desktop Mesa wrapper ($PREFIX/lib/libvulkan.so) forces compute operations into software CPU emulation or causes dispatch table collisions (SIGSEGV). termux-diffusion enforces primary binding to the host system ICD.

9.2 GPU Vendor Compatibility Matrix

• Qualcomm Adreno Series (Snapdragon):
  • Compatibility: 100% Native Production Support (Adreno 6xx, 7xx, 8xx).
  • Details: Full SPIR-V compute shader dispatch with native FP16 mixed-precision and hardware texture sampling.
• ARM Mali Series (Exynos / MediaTek Dimensity):
  • Compatibility: 100% Validated (Mali-G68, G77, G78, Immortalis).
  • Mali-Specific Mitigations: To resolve the documented Mali Bifrost/Valhall driver defect (FP16 denormal float underflow causing green image corruption or NaN divergence), termux-diffusion applies a Flush-to-Zero (FTZ) compilation pass via its verified runtime bundle (mali-compat-v2).
• Samsung Xclipse Series (AMD RDNA on Exynos 2200/2400):
  • Compatibility: Supported via Vulkan 1.3 standard compute interfaces.

10. Hardware Requirements & Operational Limits

10.1 Minimum vs. Recommended Specifications

10.2 Operational Boundaries

1. No 32-bit Support: 32-bit ARM (armv7l) environments are strictly unsupported due to address space limitations.
2. 4 GB RAM Preflight Guard: Devices equipped with only 4 GB of physical RAM must enable the --vae-tiling flag to prevent kernel OOM killer termination during VAE image reconstruction.

11. 24/7 Unattended Background Execution Guide

Follow this 3-tier hardening procedure to transform an idle Android phone into a continuous, non-throttling on-device AI generation node:

Tier 1: Termux Environment Hardening

Acquire an Android kernel CPU WakeLock to prevent the scheduler from dropping core frequencies when the display sleeps:

# Acquire permanent CPU WakeLock
termux-wake-lock

# Grant Termux read/write access to shared internal storage
termux-setup-storage

Tier 2: Android OS GUI Settings

1. Disable Battery Optimization:
   • Settings $ ightarrow$ Apps $ ightarrow$ Termux $ ightarrow$ Battery $ ightarrow$ Select 'Unrestricted'.
2. Samsung One UI Background Exemption:
   • Settings $ ightarrow$ Battery $ ightarrow$ Background usage limits $ ightarrow$ Add Termux to 'Never sleeping apps'.
3. Maximize Virtual Memory (RAM Plus):
   • Settings $ ightarrow$ Device Care $ ightarrow$ Memory $ ightarrow$ RAM Plus $ ightarrow$ Select 8 GB and reboot.

Tier 3: ADB Protocol Configuration (via USB or Wireless LADB)

Android 12 through 16 incorporate the Phantom Process Killer, which terminates background processes if total child process counts exceed 32 or sustained CPU load is detected. Completely neutralize this limitation:

# 1. Permanently disable the Android Phantom Process Killer
adb shell "/system/bin/device_config put activity_manager max_phantom_processes 2147483647"
adb shell "/system/bin/device_config set_sync_disabled_for_tests persistent"

# 2. Whitelist Termux against Android Doze deep-sleep standby
adb shell "dumpsys deviceidle whitelist +com.termux"

# 3. Lock Linux Low Memory Killer (LMK) priority (-1000 guarantees critical daemon status)
adb shell "echo -1000 > /proc/$(adb shell pidof com.termux)/oom_score_adj"

12. License & Permissible Use

termux-diffusion is released under the MIT License.

MIT License

Copyright (c) 2026 Eunho Kim (uno-km / AMEVA Foundation)

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

Summary of Rights:

• Commercial Use: Permitted freely in proprietary and commercial services.
• Modification & Distribution: Permitted with copyright notice retention.
• Private Use & Sublicensing: Permitted without royalty or attribution fees.

13. Keywords & Discoverability Index

stable-diffusion, diffusion, termux, android, samsung-galaxy, edge-ai, on-device-ai, image-generation, text-to-image, txt2img, img2img, gguf, arm64, aarch64, vulkan, vulkan-compute, spirv, adreno, mali, snapdragon, exynos, bionic-libc, taesd, vae-tiling, lora, controlnet, sdxs, sd-turbo, dreamshaper, offline-ai, mobile-inference, camera-roll, galaxy-s25, galaxy-s20

📖 Official Ecosystem Links

• AMEVA Foundation Portal: https://uno-km.vercel.app/foundation/index.html
• Interactive Web Documentation: https://uno-km.vercel.app/lib/diffusion/
• Issue Tracker: https://github.com/uno-km/termux-diffusion/issues