Z-Image Turbo Generator

Text-to-image generation CLI optimized for RTX 3060 12GB

Architecture

"a beautiful sunset"
        │
        ▼
┌───────────────────┐
│  LLM (Qwen 3 4B)  │  Text Encoder
│  qwen_3_4b-Q8_0   │  Prompt → Embedding
└───────────────────┘
        │
        ▼
┌───────────────────┐
│  Diffusion Model  │  Denoising
│  z_image_turbo    │  Noise → Latent
└───────────────────┘
        │
        ▼
┌───────────────────┐
│   VAE Decoder     │  Decoder
│   ae-f16.gguf     │  Latent → RGB
└───────────────────┘
        │
        ▼
    output.png

Requirements

• Python 3.10+
• stable-diffusion.cpp (CUDA build)
• GGUF model files

Installation

# 1. Build stable-diffusion.cpp
cd /path/to/Z-Image
git clone --recursive https://github.com/leejet/stable-diffusion.cpp
cd stable-diffusion.cpp
mkdir build && cd build
cmake .. -DSD_CUBLAS=ON
cmake --build . --config Release

# 2. Download models
cd ../gencli
bash download_zimage_files.sh

Usage

# Basic usage
python generate_zimage.py "a beautiful sunset"

# With options
python generate_zimage.py "cat on the sofa" \
    -W 768 -H 768 \      # Resolution
    --steps 8 \          # Inference steps (4-8)
    --seed 42 \          # Seed
    -n "blurry"          # Negative prompt

Options

Option	Description	Default
-W, --width	Image width	512
-H, --height	Image height	512
-s, --steps	Inference steps	8
-g, --cfg-scale	CFG scale	1.0
--seed	Random seed	random
-n, --negative	Negative prompt	-
-o, --output	Output filename	auto
-b, --batch	Batch count	1

Project Structure

gencli/
├── config.py            # Path and parameter configuration
├── generator.py         # ZImageGenerator class
├── cli.py               # CLI interface
├── generate_zimage.py   # Entry point
├── download_zimage_files.sh
├── requirements.txt
└── outputs/             # Generated images

Model Files

models/gguf/
├── z_image_turbo-Q8_0.gguf   # Diffusion (Q8 quantized)
├── ae-f16.gguf               # VAE (FP16)
└── qwen_3_4b-Q8_0.gguf       # Text Encoder (Q8 quantized)

Performance

Metric	Value
Resolution	512x512
Inference steps	4-8
Generation time	~2-4s
VRAM usage	~8-10GB

License

Educational and research purposes only.