lulavc/Z-Image-Turbo

⚡ Z-Image Turbo

Ultra-fast text-to-image generation powered by [Tongyi-MAI/Z-Image-Turbo](https://huggingface.co/Tongyi-MAI/Z-Image-Turbo) — a single-stream Diffusion Transformer that produces stunning, high-quality images in just 8 steps.

✨ Features

🖼️ Resolution Presets

Standard (1024-range)

Large (1280+)

Note: Large presets require more VRAM. If you get an out-of-memory error, switch to a Standard preset or reduce the number of steps.

🌐 Language Support

The full UI (labels, placeholders, buttons, error messages) is available in 4 languages. Switch with the language selector at the top of the control panel:

You can write prompts in any language — the model handles multilingual input natively.

✨ AI Prompt Enhancer

Click the ✨ Enhance Prompt button to automatically improve your prompt using a large language model.

What it does:

• —Locks in your core subject and intent
• —Adds professional lighting, composition, and material details
• —Applies cinematic or photographic aesthetics
• —Handles text elements with precise quotation formatting

Example:

• —Input: "a woman in a red dress"
• —Enhanced: "Cinematic close-up portrait of a woman in a flowing crimson silk dress. Soft rim lighting from the left, deep shadow on the right side of her face. Shallow depth of field, f/1.8, bokeh background of blurred warm golden lights. High fashion editorial style, Leica M10 film grain."

The enhancer uses Qwen/Qwen2.5-7B-Instruct via the HF Inference API. Generation continues normally even if enhancement is unavailable.

🎯 Quality Presets

Quick presets to balance speed and quality:

Note: Z-Image Turbo is optimized for lower guidance scales than base models. Values above 5.0 may cause oversaturation and blown highlights.

⚙️ Advanced Settings

🤖 MCP Server

This Space runs as an MCP (Model Context Protocol) server, allowing it to be used as a tool by AI agents.

Using with Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "z-image-turbo": {
      "url": "https://lulavc-z-image-turbo.hf.space/gradio_api/mcp/sse"
    }
  }
}

Using with any MCP client

• —SSE endpoint: https://lulavc-z-image-turbo.hf.space/gradio_api/mcp/sse
• —Tool name: generate
• —Parameters: prompt, negative_prompt, resolution, steps, guidance_scale, shift, seed, random_seed

🔧 Technical Details

Model Architecture

• —Name: Tongyi-MAI/Z-Image-Turbo
• —Architecture: Z-Image — single-stream Diffusion Transformer (DiT)
• —Scheduler: FlowMatch Euler Discrete with configurable time shift
• —Precision: bfloat16
• —License: Apache 2.0

Acceleration Stack

AoTI FA3 (flash-attn3 kernel)
    ↓ fallback
AoTI standard (compiled transformer blocks)
    ↓ fallback
Standard PyTorch inference

spaces.aoti_blocks_load pre-compiles the transformer blocks using Torch's Ahead-of-Time Inductor, significantly reducing per-step latency. The space automatically uses the best available kernel.

Infrastructure

• —Hardware: ZeroGPU — NVIDIA H200 (141 GB HBM3e)
• —Framework: Gradio 6.0.2
• —Scheduler cache: Shift-keyed cache avoids re-instantiation on every call
• —GPU memory: torch.cuda.empty_cache() called after every generation

💡 Tips for Best Results

1. 1.Be specific — describe lighting, style, mood, camera angle, and materials
2. 2.Name a style — "cinematic", "documentary", "concept art", "Studio Ghibli", "Leica film grain"
3. 3.Use the Enhancer — click ✨ Enhance Prompt before generating for dramatically richer output
4. 4.Start with Balanced preset — default guidance scale 1.0 provides optimal quality/speed balance
5. 5.Use Quality preset for important work — guidance scale 3.5 significantly improves detail without oversaturation
6. 6.Maximum preset for final output — guidance scale 5.0 with 10 steps provides highest quality
7. 7.Avoid high guidance scales — values above 5.0 can cause oversaturation and color artifacts in Turbo models
8. 8.Time Shift 3 — good default; try higher (5–7) for more detailed textures
9. 9.Fix your seed — uncheck Random Seed, copy the seed after a good generation, reuse it with tweaks
10. 10.Negative prompts — blur, deformed hands, watermark, low quality work well as defaults

🛠️ Running Locally

git clone https://huggingface.co/spaces/lulavc/Z-Image-Turbo
cd Z-Image-Turbo
pip install -r requirements.txt
python app.py

Requires a CUDA-capable GPU with at least 16 GB VRAM for 1024×1024. Set HF_TOKEN in your environment for prompt enhancement.

📄 License

• —Space code: MIT
• —Model weights: Apache 2.0

Space by [lulavc](https://huggingface.co/lulavc) · Powered by Tongyi-MAI/Z-Image-Turbo · ZeroGPU · A10G