Documentation

Installation

DMG Installer

Download the DMG from the releases page (~12 GB, includes model weights). Open it and drag Z-Image.app to your Applications folder.

The CLI tool is embedded in the app bundle:

# Add to your PATH (optional)
ln -s /Applications/Z-Image.app/Contents/MacOS/txt2zimage /usr/local/bin/txt2zimage

Build from Source

git clone https://gitlab.com/mlc0911-group/txt2zimage.git
cd txt2zimage
task build

Requires Go 1.21+, Task, and the MLX libraries (built automatically).

Quick Start — Desktop App

  1. Launch Z-Image.app
  2. Type a prompt in the text field
  3. Pick an aspect ratio (1:1, 4:3, 16:9, …)
  4. Click Generate
  5. Watch the step-by-step progress — your image appears in seconds

Tips for better results

  • Be descriptive: "a tabby cat sitting on a wooden table, studio lighting, shallow depth of field" works better than "cat"
  • Use the seed field to reproduce a result you like
  • Start with 9 steps (default) — increase to 12–15 for more detail
  • Higher resolutions (1536+) need 24 GB+ RAM

Quick Start — CLI

# Basic generation
txt2zimage -p "a red fox in snow" -o fox.png

# Custom resolution
txt2zimage -p "panoramic mountain view" -w 1920 -h 1080 -o panorama.png

# More steps for higher quality
txt2zimage -p "detailed portrait" -s 15 --seed 42 -o portrait.png

# Use resident server for faster generation
txt2zimage serve &          # start server in background
txt2zimage -p "prompt" -o result.png   # connects automatically

Models

Model Parameters Size Steps Description
z-image-turbo 10.3B ~12 GB (FP8) 9 Default — fast, high quality
mlczimage 10.3B + LoKr ~12 GB 9 Experimental — improved seed diversity

Models are stored in ~/.txt2zimage/models/ and downloaded automatically on first use from HuggingFace.

TeaCache

TeaCache is enabled by default. It caches intermediate denoising results and skips ~4 of 9 steps when possible — roughly 2x speedup with minimal quality impact. Disable with --no-teacache if you want maximum quality.

PNG Metadata

Every generated image embeds its generation parameters in the PNG tEXt chunks: prompt, seed, steps, resolution, model name, and TeaCache status. This means you can always look up exactly how an image was created and reproduce it.