Termux-Diffusion > AmfyUI Mobile Studio & ComfyUI DAG

AmfyUI: ComfyUI Mobile DAG Studio & Sovereign DiT Showcase

Official user manual, architecture specification, and Galaxy S20 1280×720 HD 4-Season empirical benchmarks for AmfyUI — the native Android ComfyUI DAG runtime.

🚀 Termux-Diffusion v2.0.0 Milestone: AmfyUI Production Release

AmfyUI eliminates the multi-gigabyte Python/PyTorch desktop footprint. Built directly on top of Android Bionic libc and native ARM64 NEON / Qualcomm Adreno OpenCL compute kernels, AmfyUI executes full ComfyUI JSON DAG node graphs natively on your smartphone with deterministic VRAM caps and Zero-LMK stability.

1. Architectural Nomenclature: What is AmfyUI?

AmfyUI is an engineered recursive backronym and structural design protocol tailored for sovereign edge computing:

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).

2. ComfyUI Compatibility Matrix & Operational Boundaries

AmfyUI interprets and executes official ComfyUI JSON workflows directly without running heavy desktop Python daemons.

ComfyUI Node Class AmfyUI Native Mapping Execution Backend Status
CLIPLoader / DualCLIPLoader Qwen3-4B / CLIP ViT-L / T5-XXL (GGUF / FP8) CPU NEON Multi-threading (4-8 Cores) 100% Native
UNETLoader / DiffusionModelLoader Z-Image Turbo 6.0B DiT, SDXS, SD 1.5, SDXL Adreno OpenCL 2.0 / Mali Vulkan 1.1 / CPU SDOT 100% Native
VAELoader / TAESDLoader TAESD FLUX.1 (10MB) & Standard SD VAE (FP16) CPU NEON Vectorized Decode (Zero-VRAM Spike) 100% Native
CLIPTextEncode Prompt Conditioning & Cross-Attention Vectorizer Host RAM Shared Context 100% Native
EmptyLatentImage Spatial Latent Tensor Allocator (up to 1280×720) Zero-Copy Buffer Pool 100% Native
KSampler / KSamplerAdvanced Res_Multistep, Euler, Euler_A, DPM++ 2M, LCM Tiled Flash Attention ODE Solver 100% Native
VAEDecode / VAEEncode Latent-to-RGB Reconstruction / Img2Img Tiler Spatial Tiled Decoder (--vae-tiling) 100% Native
SaveImage PNG Export & Android MediaStore Broadcaster Samsung Gallery Auto-Sync (Direct DCIM/Pictures) 100% Native
LoraLoader Dynamic Weight Additive Patching GGML Layer Merging Supported
Custom Python C++ Nodes External desktop extensions requiring GCC/CUDA Fallback to AMEVA-Cluster RPC Offloading Cluster Offload

3. Visual Node Orchestration & User Interface

AmfyUI runs an interactive web studio right inside Android Chrome or Samsung Internet. Launch the server from Termux terminal:

# 1. Launch AmfyUI Mobile Studio (Default: host=0.0.0.0, port=11553)
# Automatically binds to 0.0.0.0:11553 (both local and external LAN/Tailscale access enabled)
termux-diffusion ui

# Access URLs:
# - On phone:      http://localhost:11553
# - From PC/LAN:   http://<phone-ip>:11553 (e.g. http://100.106.99.81:11553)

# 2. Optional: Custom port & host binding:
termux-diffusion ui --port 11554 --host 0.0.0.0
AmfyUI Mobile Studio • Visual DAG Execution Flow
Device: Galaxy S20 (Snapdragon 865) | VRAM Quota: 2.0 GB
DiffusionModelLoader Model: z_image_turbo-Q8_0 Backend: GPUOpenCL VRAM Quota: 2.0 GB Max CLIPLoader (Qwen3-4B) Weights: Qwen3-4B-Q2_K Backend: CPU 4-Cores Zero VRAM Footprint CLIPTextEncode & Latent Resolution: 1280 × 720 Guidance Scale: 3.5 Tokens: 3,600 Latent Flow Shift: 3.0 KSampler (Flash Attention) Steps: 8 Iterations Sampler: res_multistep Scheduler: simple CFG Scale: 1.0 Adreno OpenCL Speedup: 2.35× NaN / Zero: 0 / 0 (Forensic) VAEDecode & SaveImage TAESD FLUX.1 (10MB) Output: 1280×720 PNG Gallery Auto-Sync: Active Bit Depth: 24-bit TrueColor

4. Galaxy S20 (Snapdragon 865) 1280×720 HD 4-Season Master Benchmark

Physical test run (Jira SCRUM-481) executed on Samsung Galaxy S20 5G (Qualcomm Snapdragon 865 KONA, 12GB LPDDR5 RAM, Adreno 650 GPU OpenCL 2.0). Generating wide 16:9 high-definition 1280×720 portrait photography across four seasonal conditions.

🔬 Key Engineering Findings from SCRUM-481
  • 2.35× Hardware Acceleration via OpenCL Flash Attention: GPU execution reduced 8-step inference latency from 9,022s (CPU baseline) down to 3,835s.
  • Cortex-A77 SDOT Hardware Advantage: On CPU, Q8_0 (8,534s) executed 488 seconds faster than Q4_0 (9,022s) because ARMv8.2-A hardware SDOT instructions execute 4 FMAs per cycle without software bit-shifting overhead.
  • Thermal Efficiency: OpenCL GPU execution operated at 41.4°C–41.8°C, which is ~6°C cooler than full CPU saturation (47.8°C).
  • Tensor Forensic Integrity: 32 individual checkpoint tensors across 4 stages recorded exactly NaN=0, Zero=0.
Season & Stage Weight Quant Execution Backend Total Elapsed Step Latency GPU Speedup Peak Thermal Forensic Sanity
🌸 1. Spring (봄) Q4_0 (3.53 GB) CPU (ARM NEON) 9,022s (2h 30m) ~1,180s / step 1.00× (Baseline) 47.8°C NaN=0 / Zero=0
☀️ 2. Summer (여름) Q4_0 (3.53 GB) Adreno OpenCL 2.0 3,835s (1h 03m) ~468s / step 2.35× Accelerated 41.4°C NaN=0 / Zero=0
🍂 3. Autumn (가을) Q8_0 (6.13 GB) CPU (ARM SDOT) 8,534s (2h 22m) ~1,060s / step 1.06× (vs CPU Q4) 47.1°C NaN=0 / Zero=0
❄️ 4. Winter (겨울) Q8_0 (6.13 GB) Adreno OpenCL 2.0 3,895s (1h 04m) ~475s / step 2.19× Accelerated 41.8°C NaN=0 / Zero=0

5. 4-Season 1280×720 HD Portrait Gallery & Prompt Engineering

Directly rendered on Samsung Galaxy S20 without upscaling. All images are 1280×720 24-bit TrueColor PNG artifacts generated using 8-step Z-Image Turbo with TAESD VAE.

Stage 1 • Spring 1280 × 720 HD Galaxy S20 Spring 1280x720 Render
Spring Sunlight • Yellow One-Piece Dress
⏱️ 9,022s (CPU) z_image_turbo Q4_0 47.8°C Mean [156.7, 147.2, 140.9]
"A stunningly beautiful 20-year-old Korean woman smiling warmly on a sunny spring day, exquisite beautiful facial features, highly detailed face, wearing an elegant yellow dress, soft spring sunlight, cherry blossom petals falling, cinematic lighting, photorealistic 8k, landscape 16:9"
📥 Download Spring Workflow (.json)
Stage 2 • Summer 1280 × 720 HD Galaxy S20 Summer 1280x720 Render
Summer Coastal • Airy Beach Dress
⏱️ 3,835s (GPU) ⚡ 2.35× OpenCL Speedup z_image_turbo Q4_0 41.4°C Cool
"A stunningly beautiful 20-year-old Korean woman smiling refreshingly on a bright summer beach, exquisite beautiful facial features, highly detailed face, wearing stylish airy summer dress, vibrant oceanic breeze, bright sunny coastal atmosphere, cinematic lighting, photorealistic 8k, landscape 16:9"
📥 Download Summer Workflow (.json)
Stage 3 • Autumn 1280 × 720 HD Galaxy S20 Autumn 1280x720 Render
Autumn Foliage • Knit Sweater
⏱️ 8,534s (CPU SDOT) ⚡ Q8 Faster than Q4 z_image_turbo Q8_0 47.1°C
"A stunningly beautiful 20-year-old Korean woman smiling peacefully in a scenic autumn park, exquisite beautiful facial features, highly detailed face, wearing a cozy autumn knit sweater, golden afternoon sunlight, colorful maple foliage, cinematic lighting, photorealistic 8k, landscape 16:9"
📥 Download Autumn Workflow (.json)
Stage 4 • Winter 1280 × 720 HD Galaxy S20 Winter 1280x720 Render
Winter Snowy • Elegant Coat & Scarf
⏱️ 3,895s (GPU) ⚡ 2.19× OpenCL Speedup z_image_turbo Q8_0 41.8°C Cool
"A stunningly beautiful 20-year-old Korean woman smiling warmly on a crisp winter day, exquisite beautiful facial features, highly detailed face, wearing an elegant winter coat and scarf, soft winter sunlight, snowy winter atmosphere, cinematic lighting, photorealistic 8k, landscape 16:9"
📥 Download Winter Workflow (.json)

6. ComfyUI Workflow JSON Download & Direct CLI Reproduction

Directly load these workflow files into AmfyUI Mobile Studio or execute them unattended via CLI:

📦 Spring 1280×720 CPU Workflow JSON (14.5 KB) 📦 Winter 1280×720 GPU OpenCL Workflow JSON (14.4 KB) 📦 Standard Z-Image Turbo Q4 DAG JSON (13.5 KB)

Executing Workflow via Terminal CLI

# Run workflow file end-to-end with real-time progress
termux-diffusion workflow run workflows/test2_s20_opencl_winter.json \
  --device opencl \
  --output-dir /sdcard/Pictures/TermuxDiffusion

# Validate and inspect workflow topology before execution
termux-diffusion workflow inspect workflows/test1_s20_cpu_spring.json

Executing Workflow via Python SDK

from termux_diffusion.workflow import WorkflowExecutor

# Execute ComfyUI DAG programmatically
executor = WorkflowExecutor()
result_image = executor.execute_file(
    "workflows/test2_s20_opencl_winter.json",
    device="opencl",
    output_dir="/sdcard/Pictures/TermuxDiffusion"
)
print(f"Generated 1280x720 TrueColor PNG: {result_image}")