Termux-Diffusion (Python)

AmfyUI (ComfyUI Mobile DAG Studio) & Sovereign Z-Image Turbo (6.0B DiT) On-Device Diffusion Acceleration for Android Termux.
Zero PRoot. Zero Virtualization. 100% Native ARM64 Bionic libc, Qualcomm Adreno OpenCL 2.0 / ARM Mali Vulkan 1.1+ GPU Shaders & NEON SIMD Vector Engine.

🚀 Major Milestone v2.0.0: AmfyUI & Sovereign 6.0B DiT Execution

termux-diffusion v2.0.0 introduces AmfyUI — an on-device, zero-dependency ComfyUI Directed Acyclic Graph (DAG) visual studio and workflow runtime designed specifically for mobile smartphone viewports. By eliminating heavy desktop Python/PyTorch dependencies (saving over 3.5GB of RAM), AmfyUI compiles and executes ComfyUI JSON workflows directly against Android Bionic libc, Qualcomm Adreno OpenCL, and ARM Mali Vulkan compute kernels.

🏛️ Architectural Nomenclature: What is AmfyUI?

AmfyUI is an engineered recursive backronym and structural design protocol:

• A — AMEVA / Asynchronous: Tri-Engine asymmetric pipelining separating Text Encoder (CPU), DiT Backbone (GPU OpenCL/Vulkan), and Latent Decoder (Host CPU).
• M — Mobile-First Memory: Strict hardware VRAM quota bounds (--max-vram) and dynamic layer streaming preventing Android Low Memory Killer (LMK) aborts.
• F — Flow-Based DAG: 100% ComfyUI UI & API Prompt JSON compatibility with topological sorting and acyclic dependency resolution.
• Y — Yield-Optimized Stream: Zero-copy mmap tensor paging and AXI bus streaming delivering maximal throughput on thermal-constrained mobile SoCs.
• UI — User Interface Studio: Ultra-lightweight zero-dependency web canvas studio optimized for touch gestures and smartphone viewports (360px–430px).

🧩 ComfyUI Compatibility Matrix & Supported Nodes

AmfyUI interprets and executes official ComfyUI JSON graph exports directly:

📊 Physical Empirical Benchmark: Galaxy S20 HD 1280×720 4-Season Showcase

Jira Ticket: SCRUM-481 | Physical Device: Samsung Galaxy S20 5G
SoC: Qualcomm Snapdragon 865 KONA (Cortex-A77 × 4 + A55 × 4)
GPU: Qualcomm Adreno 650 (OpenCL 2.0 Full Profile) | RAM: 12GB LPDDR5
Model: Tongyi Wanxiang Z-Image Turbo (6.0B DiT) + Qwen3-4B LLM + TAESD
Synthesis Resolution: 1280 × 720 (16:9 Wide High-Definition), 8 steps

🔬 Key Discoveries

• 2.35× GPU Speedup: OpenCL Flash Attention cut generation latency from 2.5 hours down to 1 hour.
• Cortex-A77 SDOT Advantage: Q8_0 (8,534s) was 488s faster than Q4_0 (9,022s) due to ARMv8.2-A hardware SDOT instructions.
• Thermal Management: Adreno OpenCL GPU ran at 41.4°C, ~6°C cooler than CPU saturation (47.8°C).
• Sanity: 32 distinct layer checkpoints verified with NaN=0, Zero=0.

💻 Python SDK Installation & Practical User Manual

1. Installation & Engine Provisioning

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

2. Python API Practical Examples

[Example 1] Programmatic ComfyUI Workflow Execution

from termux_diffusion.workflow import WorkflowExecutor

# Execute ComfyUI DAG JSON directly on Android
executor = WorkflowExecutor()
result_path = executor.execute_file(
    "workflows/test2_s20_opencl_winter.json",
    device="opencl",
    output_dir="/sdcard/Pictures/TermuxDiffusion"
)
print(f"Workflow 1280x720 TrueColor PNG saved to: {result_path}")

[Example 2] High-Level 6.0B DiT Image Synthesis

import termux_diffusion as td

# Official Z-Image Turbo 6.0B DiT Synthesis
image_path = td.generate(
    prompt="A stunningly beautiful 20-year-old Korean woman smiling warmly on a crisp winter day, 8k",
    preset="z-image-turbo",
    width=1280,
    height=720,
    steps=8,
    cfg_scale=1.0,
    guidance=3.5,
    device="opencl",
    diffusion_fa=True,
    max_vram="GPUOpenCL=2.0",
    output_path="/sdcard/Pictures/TermuxDiffusion/winter_portrait.png"
)
print(f"Generated via 6.0B DiT & Synced to Samsung Gallery: {image_path}")

[Example 3] Launching AmfyUI Mobile Server from Python

from termux_diffusion.web import run_web_server

# Start AmfyUI Mobile Studio Daemon
run_web_server(host="0.0.0.0", port=11553, blocking=True)

🌐 Official Documentation & Ecosystem Links

• Official AmfyUI Portal & Technical Report: https://uno-km.vercel.app/lib/diffusion/amfyui.html
• Spring 1280×720 CPU Workflow JSON: test1_s20_cpu_spring.json
• Winter 1280×720 OpenCL Workflow JSON: test2_s20_opencl_winter.json
• AMEVA Open-Source Foundation Portal: https://uno-km.vercel.app/foundation/index.html
• GitHub Repository: https://github.com/uno-km/termux-diffusion

📄 License

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