Z-Image: an online image generation tool based on the 6B model launched by Ali Tongyi Labs

Z-Image is an efficient image generation base model developed and open-sourced by Alibaba Tongyi Lab. It adopts an innovative architecture called Scalable Single-Stream DiT (S3-DiT) to unify text, visual semantics, and image latent variables in a single stream, thus greatly improving parameter efficiency. Unlike mega-models with tens of billions of parameters, Z-Image contains only 6 billion (6B) parameters, yet produces photo-realistic images comparable to the top commercial models. The most notable feature of the model is that it is "production friendly", with fast inference speeds (Turbo version achieves sub-second images) and low hardware requirements, running smoothly on consumer graphics cards with up to 16GB of video memory. In addition, Z-Image solves the pain point of traditional graphical models in text processing, and is able to accurately render complex Chinese and English text, which is a representative work in the open source community that balances performance, efficiency and text generation capability.

Function List

• High quality image generation: Produces photo-realistic, detailed and beautifully composed images based on 6B parametric scale.
• Bilingual text renderingThe unique text encoding processing capability enables it to accurately generate complex Chinese characters and English characters in pictures, solving the problem of "AI can't read".
• Extreme Reasoning (Turbo Mode): Provided Z-Image-Turbo version, which reduces the inference steps to 8 through distillation, enables sub-second generation on enterprise GPUs, and is extremely fast on consumer graphics cards.
• low memory footprint: The carefully optimized architecture allows the model to run on graphics cards with less than 16GB of VRAM, such as the RTX 4080/4090 or even lower memory configurations.
• Precision instructions are followed::Z-Image-Edit version is specifically fine-tuned for image editing and is able to understand complex natural language commands to make local modifications or global style conversions to images.
• Single-stream architecture (S3-DiT): The adoption of a single-stream architecture with full parameter sharing, instead of the traditional dual-stream (text-graph separation) design, enhances the depth of the model's understanding of graphical relationships.

Using Help

Z-Image offers a variety of ways to use it, both for developers through Python code calls and for designers through the ComfyUI and other visualization interfaces to use. The following are detailed operating guidelines based on general users and developers.

1. Hardware preparation

Before you begin, make sure your computer meets the following basic requirements:

• operating system: Linux or Windows (Windows 10/11 recommended).
• Graphics card (GPU): NVIDIA graphics card with 16GB or more recommended video memory (Turbo versions are optimized to run at lower video memory, but 16GB is recommended for the best experience).
• matrix: Python 3.10+ and PyTorch are installed.

2. Run with ComfyUI (recommended for designers/general users)

ComfyUI is the most popular node-based AI graph generation tool available, and Z-Image already has a community-supported workflow.

Installation Steps:

1. Download model weights::
   Visit HuggingFace or ModelScope (the Magic Hitch community), search for Z-Image-TurboThe
   Download the master model file (usually .safetensors (Format).
   Put the downloaded file into ComfyUI's models/checkpoints/ Catalog.
2. Update ComfyUI::
   Make sure your ComfyUI is up to date, or that you have installed a third-party plug-in that supports the Z-Image architecture (such as the ComfyUI-GGUF (or a specialized Z-Image loader node, depending on community updates).
3. Loading workflows::
   Download Z-Image's official or community provided workflow.json file (usually found in a GitHub repository or in the Civitai (You can find it on).
   Drag the JSON file into the ComfyUI interface.
4. Generating images::
   Enter your prompt in the "CLIP Text Encode" node. z-Image supports Chinese prompts, for example:一张海报，上面写着“通义实验室”五个大字，背景是未来的科技城市The
   Click "Queue Prompt" to start generating.

3. Run with Python code (recommended for developers)

If you are familiar with programming, you can directly use the diffusers library to run the model.

Install the dependencies:
Open a terminal or command prompt and run the following command to install the necessary libraries:

pip install torch diffusers transformers accelerate

Write run scripts:
Create a file named run_zimage.py file, fill in the following code:

import torch
from diffusers import DiffusionPipeline
# 加载 Z-Image-Turbo 模型
# 注意：如果无法直接访问 HuggingFace，请使用 ModelScope 的镜像地址
pipe = DiffusionPipeline.from_pretrained(
"Tongyi-MAI/Z-Image-Turbo",
torch_dtype=torch.bfloat16,
use_safetensors=True
)
# 启用显存优化
pipe.enable_model_cpu_offload()
# 定义提示词（支持中文）
prompt = "一只穿着宇航服的猫在月球上喝咖啡，背景有地球，照片级真实感"
# 生成图像
image = pipe(
prompt=prompt,
num_inference_steps=8,  # Turbo 版本仅需 8 步
guidance_scale=0.0      # Turbo 版本通常设为 0
).images[0]
# 保存图片
image.save("z_image_result.png")

Perform generation:
Runs in the terminal:

python run_zimage.py

At the end of the run, a file named z_image_result.png The pictures.

4. Advanced features: image editing

If you need to modify an existing image, please download Z-Image-Edit model weights and use a similar code structure, but load the Image-to-Image related Pipeline and provides an initial image as input.

application scenario

1. E-commerce poster design
   Designers can utilize Z-Image's powerful text rendering capabilities to directly generate e-commerce poster backgrounds that contain the correct product name and tagline, eliminating the need for extensive post-production PS synthesis of text and dramatically shortening the design process.
2. Social Media Content Creation
   Self-media creators can use Chinese prompts to quickly generate graphics that fit the Chinese cultural context, such as holiday greeting graphics, ancient style illustrations, etc., and don't need to worry about the complex threshold of English prompts.
3. Game asset prototyping
   Game developers can quickly iterate on concept art for game characters or scenes on a 16GB graphics memory development machine, taking advantage of the sub-second speed of the Turbo version for real-time inspired visualization.
4. Education and Documentation
   Teachers or document writers can generate diagrams or illustrations with explanatory text that utilize the model's world knowledge to accurately portray scientific phenomena or historical scenarios.

QA

1. This website address z-img.org Why won't it open?
   The URL you provided z-img.org It is highly likely that it is an old domain name that has lapsed or an address that has been misrepresented. This article describes the Z-Image The project is officially hosted primarily on GitHub (github.com/Tongyi-MAI/Z-Image) and HuggingFace platforms. Please visit these official code hosting platforms directly for resources.
2. What are the advantages of Z-Image over Stable Diffusion (SDXL)?
   Z-Image's core strengths areefficiencycap (a poem)Chinese Language Proficiency. It does this while maintaining the 6B parameters (larger than the SDXL but smaller than the Flux The S3-DiT architecture achieves very high inference speeds while being small, and natively supports Chinese cue words and Chinese text generation, which would normally require an additional ControlNet on SDXL.
3. What is the minimum amount of video memory needed to run Z-Image?
   Officially, 16GB of video memory is recommended for best performance. However, it is possible to run a card with 8GB - 12GB RAM with a quantized version (e.g. GGUF format) or with Extreme Memory Optimization (CPU Offload) turned on, but the generation speed will be slower.
4. Is it commercially available?
   Be sure to check the model's License file on the HuggingFace or GitHub pages. Generally, open source models in the Ali Tongyi family are allowed for academic research, and commercial use may require specific protocols or registration, depending on the latest official statement.