Z Image Studio

What is Z Image Studio about?

Z Image Studio provides a unified toolbox for local text‑to‑image generation powered by the Tongyi‑MAI/Z‑Image‑Turbo model and its quantized variants. It bundles three access methods—CLI, web UI, and an MCP server—so developers, artists, and AI agents can invoke image generation in the way that best fits their workflow.

How to use Z Image Studio?

1. Installation
   • Global CLI: uv tool install git+https://github.com/iconben/z-image-studio.git or pip install z-image-studio.
   • Windows: download the installer from the releases page.
2. CLI generation

   zimg generate "A futuristic city with neon lights" --width 1920 --height 1080 --steps 20 --precision q8
   

   Options include custom output path, seed, LoRA files, and disabling history.
3. Web UI

   zimg serve            # starts on http://localhost:8000
   

   Open the URL in a browser to access the responsive interface, switch themes, browse history, and manage LoRAs.
4. MCP server
   • Stdio transport for local agents: zimg-mcp (or zimg mcp).
   • When the web server runs, HTTP endpoints /mcp (streamable) and /mcp-sse (SSE) become available for remote agents.
   • Clients can call tools generate, list_models, and list_history.

Key features of Z Image Studio

• Hybrid interfaces: CLI, web UI, and MCP server in one repository.
• GPU‑aware acceleration: Automatic detection of CUDA, MPS (Apple Silicon), and ROCm; falls back to CPU.
• Attention slicing & auto‑dimension adjustment to avoid OOM errors.
• Multiple LoRA support (up to 4 simultaneous LoRAs with per‑LoRA strength).
• Seed control for reproducible results.
• Hardware‑aware model recommendation (full, q8, q4 precision) displayed in UI and CLI.
• Multilingual UI (English, Japanese, Simplified/Traditional Chinese).
• History browser with pagination and searchable metadata.
• Image sharing: download, OS‑level share, clipboard copy, and preview thumbnails.
• Theme switching (light, dark, auto) and mobile‑responsive layout.
• Configurable data & output directories via environment variables or ~/.z-image-studio/config.json.

Use cases of Z Image Studio

• Creative prototyping: Rapidly iterate on concepts for concept art, game assets, or marketing visuals.
• AI‑agent integration: Agents can request images via MCP without handling low‑level model code.
• Batch generation: Scripts can call zimg generate with different seeds or LoRAs for dataset creation.
• Cross‑platform development: Works on Windows, macOS (including Apple Silicon), and Linux with AMD GPUs.
• Educational demos: Showcases of diffusion models with a simple web interface.

FAQ from the Z Image Studio

Q: Which GPU does Z Image Studio support? A: It automatically detects NVIDIA CUDA, Apple Silicon MPS, and AMD ROCm. If none are available, it runs on CPU.

Q: How do I use LoRA models? A: Upload LoRA files through the web UI or pass --lora <file[:strength]> multiple times on the CLI (max 4). Strength values are clamped between -1.0 and 2.0.

Q: What if I get out‑of‑memory errors? A: The app enables attention slicing automatically for low‑memory systems and adjusts image dimensions to multiples of 16.

Q: Can I run the MCP server remotely? A: Yes. Start the web server (zimg serve) and use the /mcp (streamable HTTP) or /mcp-sse endpoints. Clients should try /mcp first and fall back to SSE if needed.

Q: How to enable torch.compile on Python 3.12? A: Set the environment variable ZIMAGE_ENABLE_TORCH_COMPILE=1 or add it to ~/.z-image-studio/config.json. This is experimental and may cause errors.

Q: Where are generated images stored? A: By default, under ~/.local/share/z-image-studio/outputs (Linux), ~/Library/Application Support/z-image-studio/outputs (macOS), or %LOCALAPPDATA%\z-image-studio\outputs (Windows). Paths can be overridden with Z_IMAGE_STUDIO_OUTPUT_DIR.

Z-Image Studio

A Cli, a webUI, and a MCP server for the Z-Image-Turbo text-to-image generation model (Tongyi-MAI/Z-Image-Turbo and its variants).

This tool is designed to run efficiently on local machines for Windows/Mac/Linux users. It features specific optimizations for NVIDIA (CUDA), Apple Silicon (MPS), and AMD on Linux (ROCm), falling back to CPU if no compatible GPU is detected.

Features

Hybrid Interfaces:

• CLI: Fast, direct image generation from the terminal.
• Web UI: Modern web interface for interactive generation.
• MCP Server: Capability to be called by AI agents.

CLI and core features

• Z-Image-Turbo Model: Utilizes the high-quality Tongyi-MAI/Z-Image-Turbo model and quatized variants via diffusers.
• MPS Acceleration: Optimized for Mac users with Apple Silicon.
• ROCm Support: Explicitly supported on Linux for AMD GPUs.
• Attention Slicing Auto-detection: Automatically manages memory usage (e.g., enables attention slicing for systems with lower RAM/VRAM) to prevent Out-of-Memory errors and optimize performance.
• Seed Control: Reproducible image generation via CLI or Web UI.
• Multiple LoRA Support: Upload/manage LoRAs in the web UI, apply up to 4 with per-LoRA strengths in a single generation; CLI supports multiple --lora entries with optional strengths.
• Automatic Dimension Adjustment: Ensures image dimensions are compatible (multiples of 16).
• Customizable Output Directory: Image output directory can be customized via config file and environment variable.

Web UI features

• Multilanguage Support: English, Japanese, Chinese Simplified (zh-CN), and Chinese Traditional (zh-TW) are supported.
• History Browser: Efficiently search and browse your past generations with a paginated history that loads more items as you scroll.
• Hardware-aware Model Recommendation: The Web UI dynamically presents model precision options based on your system's detected RAM/VRAM, recommending the optimal choice for your hardware. You can also inspect available models and recommendations via the CLI.
• Image Sharing: The generated image can be downloaded to browser download directory, conveniently shared via OS share protocol, and copied into clipboard.
• Theme Switch: Light, dark and auto themes.
• Mobile compatible: Responsive layout for mobile devices.

MCP features

• MCP Server (stdio + SSE + Streamable HTTP): Expose tools for image generation, listing models, and viewing history over Model Context Protocol; stdio entrypoints (zimg mcp, zimg-mcp) for local agents, SSE available at /mcp-sse, and MCP 2025-03-26 Streamable HTTP transport at /mcp.
• Transport-Agnostic Content: All transports (stdio, SSE, Streamable HTTP) return identical structured content for consistent agent integration.
• Client Transport Selection: Clients should try Streamable HTTP (/mcp) first for optimal performance, falling back to SSE (/mcp-sse) if needed.

Requirements

• Python >= 3.11
• uv (recommended for dependency management)

Python 3.12+ Note: torch.compile is disabled by default for Python 3.12+ due to known compatibility issues with the Z-Image model architecture. If you want to experiment with torch.compile on Python 3.12+, set ZIMAGE_ENABLE_TORCH_COMPILE=1 via environment variable or in ~/.z-image-studio/config.json (experimental, may cause errors).

GPU acceleration notes

• NVIDIA (CUDA): Works with standard PyTorch CUDA builds.
• Apple Silicon (MPS): Uses PyTorch MPS backend on macOS.
• AMD on Linux (ROCm): Explicitly supported on Linux.

  Note: AMD GPU support currently requires ROCm, which is only available for Linux PyTorch builds. Windows users with AMD GPUs will currently fall back to CPU.

  • Installation: Install AMD ROCm drivers/runtime for your distribution. Then install PyTorch with ROCm support (e.g., via pip install torch --index-url https://download.pytorch.org/whl/rocm6.1 or similar). Ensure the PyTorch ROCm version matches your installed driver version.
  • Verification: The app will automatically detect your device as "rocm". You can confirm this by running zimg models.
  • Troubleshooting:
    • If the app falls back to CPU, ensure torch.version.hip is detected.
    • HSA Override: For some consumer GPUs (e.g., RX 6000/7000 series) not officially supported by all ROCm versions, you may need to set HSA_OVERRIDE_GFX_VERSION (e.g., 10.3.0 for RDNA2, 11.0.0 for RDNA3).
    • Performance: torch.compile is disabled by default on ROCm due to experimental support. You can force-enable it with ZIMAGE_ENABLE_TORCH_COMPILE=1 if your setup (Triton/ROCm version) supports it.

Global installation

If you just want the zimg CLI to be available from anywhere, install it as a uv tool:

uv tool install git+https://github.com/iconben/z-image-studio.git
# or, if you have the repo cloned locally:
# git clone https://github.com/iconben/z-image-studio.git
# cd z-image-studio
# uv tool install .

After this, the zimg command is available globally:

zimg --help

To update z-image-studio:

uv tool upgrade z-image-studio
# or, if you have the repo cloned locally, you pull the latest source code:
# git pull

Windows Installation

For Windows users, a pre-built installer is available that bundles everything you need:

1. Download the latest installer from GitHub Releases
2. Run Z-Image-Studio-Setup-x.x.x.exe
3. Follow the installation wizard
4. Launch from the Start Menu:
   • Z-Image Studio (Web UI): Starts the web server and opens your browser
   • Z-Image Studio CLI: Opens a console for command-line usage

Installation Details

• Install Location: C:\Program Files\Z-Image Studio
• User Data: %LOCALAPPDATA%\z-image-studio (contains database, LoRAs, and outputs)
• Uninstall: Use "Add or Remove Programs" or the uninstall shortcut in the Start Menu

System Requirements

• Windows 10 or Windows 11
• NVIDIA GPU with CUDA support (recommended) or compatible AMD GPU
• 8GB+ RAM (16GB+ recommended for full precision models)

pip / uv Installation

Install Z-Image Studio via pip or uv:

pip install z-image-studio
# or
uv pip install z-image-studio

After installation, the zimg command is available globally:

zimg --help

From Source

git clone https://github.com/iconben/z-image-studio.git
cd z-image-studio
pip install -e .
# or
uv pip install -e .

Usage

After installation, you can use the zimg command directly from your terminal.

1. CLI Generation (Default Mode)

Generate images directly from the command line using the generate (or gen) subcommand.

# Basic generation
zimg generate "A futuristic city with neon lights"

# Using the alias 'gen'
zimg gen "A cute cat"

# Custom output path
zimg gen "A cute cat" --output "my_cat.png"

# High quality settings
zimg gen "Landscape view" --width 1920 --height 1080 --steps 20

# With a specific seed for reproducibility
zimg gen "A majestic dragon" --seed 12345

# Select model precision (full, q8, q4)
zimg gen "A futuristic city" --precision q8

# Skip writing to history DB
zimg gen "Quick scratch" --no-history

2. Web Server Mode

Launch the web interface to generate images interactively.

# Start server on default port (http://localhost:8000)
zimg serve

# Start on custom host/port
zimg serve --host 0.0.0.0 --port 9090

Once started, open your browser to the displayed URL.

3. MCP Server Mode (Model Context Protocol)

Run Z-Image Studio as an MCP server:

# stdio transport (ideal for local agents/tools); also available as `zimg mcp`
zimg-mcp

# MCP transports are available when you run the web server:
zimg serve          # Both Streamable HTTP (/mcp) and SSE (/mcp-sse) available
zimg serve --disable-mcp   # Disable all MCP endpoints

Available tools: generate (prompt to image), list_models, and list_history. Logs are routed to stderr to keep MCP stdio clean.

Connecting an AI agent (e.g., Claude Desktop) to zimg-mcp

1. Ensure dependencies are installed (uv sync) and that zimg-mcp is on PATH (installed via uv tool install . or run locally via uv run zimg-mcp).

2. In Claude Desktop (or any MCP-aware client), add a local mcp server entry like:

   {
     "mcpServers": {
       "z-image-studio": {
         "command": "zimg-mcp",
         "args": [],
         "env": {}
       }
     }
   }
   

   Adjust the command to a full path if not on PATH. If the agent cannot find the zimg-mcp command, you can also try setting the path in environment.

   Different agents may have slightly different parameters, for example, cline will timeout fast if you do not explicitly set a timeout parameter. Here is the example for cline:

   {
     "mcpServers": {
       "z-image-studio": {
         "command": "zimg-mcp",
         "type": "stdio",
         "args": [],,
         "env": {},
         "disabled": false,
         "autoApprove": [],
         "timeout": 300
       }
     }
   }
   

   Detailed syntax may vary, please refer to the specific agent's documentation.

3. For Clients that support remote mcp server, configure the client with the streamable Http mcp endpoint URL (meanwhile keep the server up by running zimg serve). Here is an example for Gemini CLI:

   {
     "mcpServers": {
       "z-image-studio": {
         "httpUrl": "http://localhost:8000/mcp"
       }
     }
   }
   

   Detailed syntax may vary, please refer to the specific agent's documentation.

4. For legacy SSE , run zimg serve and configure the client with the SSE endpoint URL. Here is an example for Cline CLI:

   {
     "mcpServers": {
       "z-image-studio": {
         "url": "http://localhost:8000/mcp-sse/sse"
       }
     }
   }
   

   Detailed syntax may vary, please refer to the specific agent's documentation.

5. The agent will receive tools: generate, list_models, list_history.

MCP Content Structure

The generate tool returns a consistent content array with three items in this order:

1. TextContent: Enhanced metadata including generation info, file details, and preview metadata

   {
     "message": "Image generated successfully",
     "duration_seconds": 1.23,
     "width": 1280,
     "height": 720,
     "precision": "q8",
     "model_id": "z-image-turbo-q8",
     "seed": 12345,
     "filename": "image_12345.png",
     "file_path": "/absolute/path/to/image_12345.png",
     "access_note": "Access full image via ResourceLink.uri or this URL",
     "preview": true,
     "preview_size": 400,
     "preview_mime": "image/png"
   }
   

   • SSE/Streamable HTTP Transports: url and access_note point to the absolute image URL
   • Stdio Transport: file_path and access_note point to the local file path

2. ResourceLink: Main image file reference with context-appropriate URI

   • SSE/Streamable HTTP Transports: Absolute URL built from request context, ZIMAGE_BASE_URL, or relative path
   • Stdio Transport: file:// URI for local access

   URI Building Priority (SSE/Streamable HTTP):

   1. Request Context (via Context parameter) - builds absolute URL from X-Forwarded-* headers
   2. ZIMAGE_BASE_URL environment variable - configured base URL
   3. Relative URL - fallback when no other method available

   Example with Context parameter:

   @mcp.tool()
   async def generate_with_context(..., ctx: Context) -> ...:
       request = ctx.request_context.request
       proto = request.headers.get('x-forwarded-proto', 'http')
       host = request.headers.get('x-forwarded-host', 'localhost')
       return ResourceLink(uri=f"{proto}://{host}/outputs/image.png", ...)
   

   {
     "type": "resource_link",
     "name": "image_12345.png",
     "uri": "https://example.com/outputs/image_12345.png",
     "mimeType": "image/png"
   }
   

3. ImageContent: Thumbnail preview (base64 PNG, max 400px)

   {
     "data": "base64-encoded-png-data",
     "mimeType": "image/png"
   }
   

This structure ensures:

• ✅ Consistency: Same content for both stdio and SSE transports
• ✅ Efficiency: No URL/path duplication across content items
• ✅ Flexibility: ResourceLink provides file access while ImageContent offers immediate preview
• ✅ Compatibility: Follows MCP best practices for structured content types

Command Line Arguments

Subcommand: generate (alias: gen)

Argument	Short	Type	Default	Description
prompt		str	Required	The text prompt for image generation.
--output	-o	str	None	Custom output filename. Defaults to outputs/<prompt-slug>.png inside the data directory.
--steps		int	9	Number of inference steps. Higher usually means better quality.
--width	-w	int	1280	Image width (automatically adjusted to be a multiple of 8).
--height	-H	int	720	Image height (automatically adjusted to be a multiple of 8).
--seed		int	None	Random seed for reproducible generation.
--precision		str	q8	Model precision (full, q8, q4). q8 is the default and balanced, full is higher quality but slower, q4 is fastest and uses less memory.
--lora		str	[]	LoRA filename or path, optionally with strength (name.safetensors:0.8). Can be passed multiple times (max 4); strength is clamped to -1.0..2.0.
--no-history		bool	False	Do not record this generation in the history database.

Subcommand: serve

Argument	Type	Default	Description
--host	str	0.0.0.0	Host to bind the server to.
--port	int	8000	Port to bind the server to.
--reload	bool	False	Enable auto-reload (for development).
--timeout-graceful-shutdown	int	5	Seconds to wait for graceful shutdown before forcing exit.
--disable-mcp	bool	False	Disable all MCP endpoints (/mcp and /mcp-sse).

Subcommand: models

Argument	Short	Type	Default	Description
(None)				Lists available image generation models, highlights the one recommended for your system's hardware, and displays their corresponding Hugging Face model IDs.

Subcommand: mcp

Argument	Type	Default	Description
(none)			Stdio-only MCP server (for agents). Use zimg-mcp or zimg mcp.

Data Directory and Configuration

By default, Z-Image Studio uses the following directories:

• Data Directory (Database, LoRAs): ~/.local/share/z-image-studio (Linux), ~/Library/Application Support/z-image-studio (macOS), or %LOCALAPPDATA%\z-image-studio (Windows).
• Output Directory (Generated Images): <Data Directory>/outputs by default.

Configure the directory

• Config File: ~/.z-image-studio/config.json (created on first run after migration).
  • Override the data directory with Z_IMAGE_STUDIO_DATA_DIR.
  • If you want the output directory sit in another location instead of the data directory, you can override it with Z_IMAGE_STUDIO_OUTPUT_DIR.

Directory structure inside Data Directory by default:

• zimage.db: SQLite database
• loras/: LoRA models
• outputs/: Generated image files

One-time Migration (automatic)

On first run without an existing config file, the app migrates legacy data by moving:

• outputs/, loras/, and zimage.db from the current working directory (old layout) into the new locations.

Screenshots

(Screenshot 1: Two column layout with History browser collapsed)

(Screenshot 2: Three column layout with History browser pinned)

(Screenshot 3: Generated Image zoomed to fit the screen)

Development

Installation in Project Virtual Environment

1. Clone the repository:

   git clone https://github.com/iconben/z-image-studio.git
   cd z-image-studio
   

To run the source code directly without installation:

1. Run CLI:

   uv run src/zimage/cli.py generate "A prompt"
   

2. Run Server:

   uv run src/zimage/cli.py serve --reload
   

3. Run tests:

   uv run pytest
   

Optional: Install in editable mode:**

First install it:
```bash
uv pip install -e .
```

After this, the `zimg` command is available **inside this virtual environment**:

Then use the zimg command in either ways:

Using `uv` (recommended):
```bash
uv run zimg generate "A prompt"
```

or use in more traditional way:
```bash
source .venv/bin/activate  # Under Windows: .venv\Scripts\activate
zimg serve
```

Optional: Override the folder settings with environment variables

If you do not want your development data mess up your production data,
You can define environment variable Z_IMAGE_STUDIO_DATA_DIR to change the data folder for
You can also define environment variable Z_IMAGE_STUDIO_OUTPUT_DIR to change the output folder to another separate folder

Environment Variables

Variable	Description
ZIMAGE_ENABLE_TORCH_COMPILE	Force enable torch.compile optimization (experimental). By default disabled for Python 3.12+ due to known compatibility issues. Can be set to 1 via environment variable or config file (~/.z-image-studio/config.json) to enable at your own risk.
Z_IMAGE_STUDIO_DATA_DIR	Override the default data directory location.
Z_IMAGE_STUDIO_OUTPUT_DIR	Override the default output directory location.

Notes

• Guidance Scale: The script hardcodes guidance_scale=0.0 as required by the Turbo model distillation process.
• Safety Checker: Disabled by default to prevent false positives and potential black image outputs during local testing.

For detailed architecture and development guidelines, see docs/architecture.md.