Termux-Diffusion

Production On-Device AI Image Generation Framework for Android Termux & Samsung Galaxy
Dual-Engine Architecture (Python & Node.js / TypeScript) with Native Bionic ARM64 Tensor Acceleration

📖 Official Documentation Site (13 Languages) • 🌌 AMEVA Foundation • ⚡ Quickstart • 🏛️ Architecture • 📊 Benchmarks

---

1. Quick Scenario Playbook

[Install] Scenario 1: Clean Install (Fresh Setup on Android Termux)

Open the Termux application and execute the 2 commands for your preferred runtime:

[Python] Python:

# 1. Grant Android Storage Permission (Tap 'Allow' on the Android OS prompt)
termux-setup-storage

# 2. Install Toolchains & Provision Native Engine
pkg update -y && pkg install python clang make cmake git termux-api wget vulkan-loader vulkan-headers vulkan-tools -y
pip install termux-diffusion && termux-diffusion-install

[Node.js] Node.js / TypeScript:

# 1. Grant Android Storage Permission
termux-setup-storage

# 2. Install Toolchains & Provision Native Engine
pkg update -y && pkg install nodejs-lts clang make cmake git termux-api wget vulkan-loader vulkan-headers vulkan-tools -y
npm install -g termux-diffusion && npx termux-diffusion install

---

[Instant] Scenario 2: Instant Generation (Ready to Run)

Option A: One-Line CLI Generation (No Coding Required)

• Python CLI:

  termux-diffusion generate "RAW photo, portrait of happy Korean developer wearing hoodie, photorealistic" -m realistic
  

• Node.js CLI:

  npx termux-diffusion generate "RAW photo, portrait of happy Korean developer wearing hoodie, photorealistic" -m realistic
  

  (Output images are automatically synchronized to Samsung Gallery / Google Photos upon completion).

Option B: Programmatic SDK Integration

• Python (generate_demo.py):

  from termux_diffusion import generate
  
  result = generate("cyberpunk cat with neon collar in rainy alley", model="speed")
  print(f"Output saved to: {result.path}")
  print(f"Samsung Gallery Path: {result.gallery_path}")
  

• Node.js (generate_demo.js):

  const { generate } = require('termux-diffusion');
  
  async function main() {
    const result = await generate({
      prompt: 'cyberpunk cat with neon collar in rainy alley',
      model: 'speed'
    });
    console.log('Output Path:', result.path);
    console.log('Gallery Path:', result.galleryPath);
  }
  main();
  

---

[Models] Scenario 3: Custom Models & External Weights

Case A: Direct Hugging Face Repository

Provide any Hugging Face repo ID and .gguf file path. The engine auto-streams, caches, and runs it:

from termux_diffusion import generate

generate(
    "1girl, anime masterpiece, vibrant colors",
    model="second-state/DreamShaper-8-GGUF/dreamshaper-8-Q4_k.gguf"
)

Case B: Local File Reference (Internal Storage or SD Card)

generate(
    "beautiful fantasy castle at sunrise",
    model="~/storage/downloads/my_custom_model.gguf"
)

Case C: Model Aliasing (register_model)

from termux_diffusion import register_model, generate

# Register alias once
register_model("waifu", repo_id="second-state/DreamShaper-8-GGUF", filename="dreamshaper-8-Q4_k.gguf")

# Invoke cleanly anytime
generate("magical forest with fairies", model="waifu")

[Engine] Hardware Acceleration Target (device):

# Offload compute to mobile GPU (Adreno / Samsung Xclipse)
generate("speedy race car", model="speed", device="gpu")

---

2. Built-in Model Hub Presets & Golden Parameter Matrix

Each model architecture has distinct mathematical requirements for denoising steps, CFG scale, and samplers:

Preset	Actual Model Checkpoint & Quantization	Architecture Type	Optimal Steps	Optimal CFG	Recommended Sampler & Scheduler	Workload & Visual Output
"sdxs"	sdxs-512-tinySDdistilled_Q8_0 (651 MB)	1-Step Distilled	1 ~ 2	1.0	euler_a (default)	Ultra-Fast Mobile (1-2s): Clean & crisp, zero noise
"anime"	DreamShaper8_LCM_q4_0 (1.55 GB)	LCM 4~8-Step	4 ~ 8	1.5 ~ 2.0	lcm (default / karras)	Sharp 2D/2.5D Anime: Clear line art & rich cel-shading
"realistic"	realisticVisionV60B1_v51HyperVAE_Q4_k (1.55 GB)	Full SD1.5 Photoreal	20 ~ 25	6.5 ~ 7.5	dpm2 / euler_a (karras)	Ultra-Detailed Realism: Skin pores, realistic eyes, cinematic
"speed"	stable-diffusion-v1-5-Q4_1 (1.68 GB)	SD1.5 Base Q4_1	15 ~ 20	6.0 ~ 7.0	euler_a / dpm++2m (karras)	General Drafting: Balanced speed & composition fidelity
"turbo"	stable-diffusion-v1-5-pruned-emaonly_Q4_0 (1.49 GB)	SD1.5 Base Pruned	15 ~ 20	6.0 ~ 7.0	euler_a / dpm++2m (karras)	Lightweight SD1.5: Fast base generation

⚠️ Golden Rule for Distilled Models (sdxs, turbo ADD, anime LCM):
Never use high CFG ($> 2.0$) or 2nd-order ODE samplers (dpm2, heun) on distilled 1~4 step models. Doing so breaks the compressed latent manifold and causes color blowout or over-smoothing blur. Keep cfg_scale=1.0 with euler_a for crisp clarity!

💡 Golden Rule for Full SD1.5 Models (realistic, speed, turbo):
Full SD1.5 models require at least 15~20 steps with CFG=6.0~7.5 and quality-guard negative prompts to fully resolve high-frequency photorealistic details. Running them at 2~4 steps results in un-denoised noise.

---

3. Comprehensive 100% Function & API Reference

3.1 generate(...) - Primary Synthesis Function

Full parameter specification for Python and Node.js:

from termux_diffusion import generate

result = generate(
    prompt="A majestic lion standing on a cliff at sunset, 8k, photorealistic",
    negative_prompt="blurry, low quality, distorted, bad anatomy",
    model="realistic",          # Preset name ('realistic', 'speed', 'sdxs', 'turbo', 'anime'), HuggingFace ID, or local file
    device="cpu",               # 'cpu', 'gpu', 'opencl', or 'vulkan'
    output="lion.png",          # Custom destination file path (default: output_<timestamp>.png)
    width=512,                  # Image width in pixels (multiple of 64, default: 512)
    height=512,                 # Image height in pixels (multiple of 64, default: 512)
    steps=10,                   # Sampling steps (default: 10, sdxs: 2, turbo: 8)
    cfg_scale=4.0,              # Classifier-Free Guidance scale (default: 4.0)
    seed=-1,                    # Random seed (-1 for randomized generation)
    threads=4,                  # Core allocation (default: auto-detected big cores)
    wake_lock=True,             # Hold Android CPU WakeLock to prevent sleep
    export_gallery=True,        # Sync to ~/storage/pictures/TermuxDiffusion & trigger MediaScanner
    timeout=1800,               # Max execution timeout in seconds
    auto_provision=False        # Auto-compile C++ engine if missing
)

# Return Object: GenerationResult
print("Local Path:", result.path)
print("Gallery Path:", result.gallery_path)

// Node.js Equivalent
const { generate } = require('termux-diffusion');

const result = await generate({
  prompt: 'A majestic lion standing on a cliff at sunset, 8k, photorealistic',
  negativePrompt: 'blurry, low quality, distorted',
  model: 'realistic',
  device: 'cpu',
  output: 'lion.png',
  width: 512,
  height: 512,
  steps: 10,
  cfgScale: 4.0,
  seed: -1,
  threads: 4,
  wakeLock: true,
  exportGallery: true,
  timeout: 1800
});

3.2 High-Precision Controls & Advanced Parameters

termux-diffusion exposes full native C++ controls for fine-grained generation. For an exhaustive guide and all valid option lists, see ADVANCED_PARAMETERS.md.

# 1. Advanced Sampler & Noise Schedule (dpm++2m + karras photorealism)
generate(
    "hyperrealistic portrait of a cyberpunk hacker, 8k",
    sampling_method="dpm++2m",
    schedule="karras",
    steps=12
)

# 2. VAE Tiling (Reduces peak RAM by ~70% on mobile devices)
generate("epic mountain landscape", width=768, height=768, vae_tiling=True)

# 3. Image-to-Image (Img2Img Transformation)
generate(
    "convert sketch into an oil painting of a castle",
    init_img="/sdcard/Pictures/my_sketch.png",
    strength=0.70
)

# 4. LoRA Adapter Weights Injection
generate(
    "cyberpunk warrior in battle armor <lora:cyber_armor:0.8>",
    lora_dir="/data/data/com.termux/files/home/loras"
)

# 5. CLIP Skip (Anime / DreamShaper Optimization)
generate("1girl, anime masterpiece, starry night", model="anime", clip_skip=2)

# 6. ControlNet Spatial Guidance
generate(
    "warrior posing heroically",
    control_net="~/models/cnet_openpose.gguf",
    control_image="~/pose_guide.png",
    control_strength=0.9
)

3.3 Model Cache & Storage Management APIs

from termux_diffusion import (
    set_cache_dir,       # Route cache to external storage / SD card
    get_cache_dir,       # Inspect active cache directory
    download_model,      # Pre-download models in background with progress
    register_model,      # Register custom Hugging Face model presets
    list_cached_models,  # List all downloaded .gguf files and sizes
    clear_cache          # Delete cached weights to reclaim storage
)

# 1. Configure custom cache path (e.g. SD Card)
set_cache_dir("~/storage/external-1/ai_models")

# 2. Pre-fetch weights with real-time streaming progress
download_model("sdxs", force=False)

# 3. Register custom alias
register_model(
    name="cyber-waifu",
    repo_id="second-state/DreamShaper-8-GGUF",
    filename="dreamshaper-8-Q4_k.gguf",
    description="DreamShaper 8 Q4_K model for stylized anime portraits"
)

# 4. Inspect downloaded models
cached = list_cached_models()
for item in cached:
    print(f"Model: {item['name']}, Size: {item['size_mb']:.1f}MB, Path: {item['path']}")

# 5. Purge model cache
# clear_cache()

---

3.3 Hardware Diagnostics & Platform Inspection APIs

from termux_diffusion import (
    get_memory_info,            # RAM and zRAM (Samsung RAM Plus) stats
    get_optimal_thread_count,   # Optimal CPU thread affinity count
    is_android_termux,          # True if running inside Android Termux
    run_doctor,                 # Automated 7-tier system diagnostic health check
    export_to_android_gallery,  # Manually broadcast any image to Android Gallery
    TermuxWakeLock              # Context manager for holding CPU WakeLock
)

# Inspect memory safety
mem = get_memory_info()
print(f"Total RAM: {mem.total_mb}MB, Free: {mem.free_mb}MB, Swap: {mem.swap_total_mb}MB")

# Run full system diagnostic
report = run_doctor()
print(f"Doctor Health Status: {'PASSED' if report.is_ready else 'FAILED'}")

---

4. CLI Command Reference Manual

Command	Arguments	Description
termux-diffusion generate	"<prompt>" [-m model] [--device cpu|gpu] [-s steps] [-c cfg] [-t threads] [--sampler name] [--schedule name] [--vae-tiling] [-i img.png] [--strength 0.75] [--lora-dir dir] [--clip-skip 2] [--control-net cnet.gguf] [--control-image img.png] [--taesd taesd.gguf] [--seed N] [-o file.png]	Executes diffusion inference with custom options
termux-diffusion download	<model_name>	Pre-downloads and caches model weights
termux-diffusion models	(None)	Displays catalog of available presets and cached models
termux-diffusion doctor	(None)	Runs automated 7-tier pre-flight diagnostic health check
termux-diffusion install	[--force]	Compiles native ARM64 Bionic engine binary
termux-diffusion clear-cache	(None)	Clears cached weights to free storage

---

5. [Memory Optimization] Samsung RAM Plus & Low-Memory (RAM) Devices

For Android devices with 4GB - 8GB physical RAM, image synthesis models (1.5GB tensor graphs) require additional virtual swap space to prevent the Android Low Memory Killer (LMK) from terminating background tasks.

1. Enable Samsung RAM Plus (One UI 4 / 5 / 6)

1. Open Settings -> Battery and device care (or Device Care).
2. Tap Memory -> RAM Plus.
3. Select +4 GB, +6 GB, or +8 GB and restart your phone.
4. This expands available virtual memory (zRAM swap) to ensure Stable Diffusion runs seamlessly without memory pressure.

2. General Android / Non-Samsung Devices (zRAM Activation)

If your device does not have manufacturer RAM Plus, ensure zRAM swap is active:

# Verify active swap memory in Termux:
free -m

---

6. [Security] Android 12 / 13 / 14+ Phantom Process Killer Prevention

On Android 12 (API 31) and higher, the OS may kill background child processes (sd-cli) if the total process limit exceeds 32. Follow these recommended settings for 24/7 background stability:

1. Developer Options Configuration (Android 12L / 13 / 14+)

1. Open Settings -> About phone -> Software information -> Tap Build number 7 times to unlock Developer Options.
2. Go to Settings -> Developer options.
3. Enable Disable child process restrictions.

2. ADB One-Time Permanent Bypass (Optional via PC)

Connect phone to PC via USB debugging and run once:

adb shell "/system/bin/device_config set_sync_disabled_for_tests persistent"
adb shell "/system/bin/device_config put activity_manager max_phantom_processes 2147483647"

3. Battery Optimization & Background Notification

• Unrestricted Battery: Open Settings -> Apps -> Termux -> Battery -> Select Unrestricted (Don't optimize).
• WakeLock Notification: Keep the Termux notification active and tap Acquire wakelock from the status bar dropdown.

---

7. Architecture & Security Isolation

• Zero PRoot / Zero Root: Executes directly against native Android Bionic libc with ARM64 NEON SIMD optimizations, avoiding virtual container memory amplification.
• Zero Deception & Honest Diagnostics: Zero fake logs. NPU/GPU/CPU hardware is probed transparently without deceptive rerouting.
• Configurable Negative Prompt: Negative prompt defaults to None with zero bias against subjects, configurable per-call or globally.
• Process Reaper: Intercepts SIGINT / SIGTERM / KeyboardInterrupt to forcefully clean up orphaned child sd-cli processes.
• WakeLock Shield: Automatically prevents CPU sleep states when the smartphone screen turns off during lengthy inference.

---

8. The AMEVA Mobile AI & Automation Ecosystem

• Termux-Playwright (PyPI | npm | Official Docs): Production headless Chromium browser automation for Android Termux.
  • Python: pip install termux-playwright && termux-playwright-install
  • Node.js: npm install -g termux-playwright && npx termux-playwright install

---

9. Disclaimer (면책 조항)

Disclaimer:
Termux-Diffusion is an independent open-source project developed for the Android Termux environment and is not officially affiliated with, endorsed by, or sponsored by the Termux project.

(본 프로젝트는 안드로이드 Termux 환경을 위해 개발된 독립적인 오픈소스 라이브러리이며, Termux 공식 프로젝트와 직접적인 제휴 관계가 아닙니다.)

---

10. License

Released under the MIT License. Maintained by uno-km (Eunho Kim).

---

💖 Sponsorship & Community Backing

AMEVA is an independent open-source public good governed under the AMEVA Open-Source Foundation (AOSF). All sponsorship funds are 100% publicly audited and dedicated to physical ARM64 testbeds and CI/CD GPU runners.

• Open Collective (Non-Profit 501(c)(6)): https://opencollective.com/ameva-fund
• GitHub Sponsors: https://github.com/sponsors/uno-km
• Official Foundation Portal: https://uno-km.vercel.app/docs/foundation/sponsorship.html