---
description: Generate images from text prompts using local GPU inference
allowed-tools: Bash(z-image:*)
---

# Z-Image - Local AI Image Generation

## Tool

**Binary:** `z-image` (in PATH via `~/bin/z-image`)
**Script:** `~/.claude/skills/z-image/generate.py`
**Model:** Tongyi-MAI/Z-Image-Turbo (diffusers, bfloat16, CUDA)
**venv:** `~/.claude/skills/z-image/.venv/`

## Usage

```bash
# Basic generation
z-image "a cat sitting on a cloud"

# Custom output filename
z-image "sunset over mountains" -o sunset.png

# Custom output directory
z-image "forest path" -d ~/Pictures/ai-generated/

# More inference steps (higher quality, slower)
z-image "detailed portrait" -s 20

# Disable CPU offloading (faster if VRAM allows)
z-image "quick sketch" --no-offload
```

## Defaults
- **Steps:** 9 (fast turbo mode)
- **Guidance scale:** 0.0 (turbo model doesn't need guidance)
- **Output:** `zimage_TIMESTAMP_PROMPT.png` in current directory
- **VRAM:** Uses CPU offloading by default to reduce VRAM usage

## Notes
- First run downloads the model (~several GB)
- Requires NVIDIA GPU with CUDA support
- Output is always PNG format
- After generating, use the Read tool to show the image to the user