Z-Image

Z-Image

Generates images from a text description with a 6.2B-parameter model, the full undistilled base of the Z-Image family. Built for creators, researchers and developers who want close control over results.

Made by Tongyi-MAI on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Z-Image is a text to image model from Tongyi-MAI. You write a description and it returns a picture. It runs through the diffusers library and has 6.2B parameters. This is the full base model of the Z-Image family, not the speed-tuned Turbo version, so it suits people who want close control or a starting point for their own training.

Who it suits: Illustrators, researchers and developers who want an image model they can run and adapt on their own computer.

Industry | Who uses it | How they use it | Example
Art and design | Illustrators and concept artists | Describe a scene in a photo, anime or illustrated style | A character portrait in anime style
Machine learning | Researchers | Use the undistilled model as a base for LoRA or ControlNet training | A LoRA trained for one visual style
Software | Developers | Call the model from Python code through diffusers | A script that saves a picture for each prompt
Marketing | Content makers | Draft visuals from a written brief | A lifestyle photo concept for a campaign

Anyone who only wants the fastest results, at eight steps, can skip it and look at Z-Image-Turbo instead.

Where to use it:

• Windows, macOS or Linux with Python, as the download steps describe
• Any Python script or notebook that uses the diffusers library
• A browser, through the online demo the model card points to

Use case | Who | What they get back
Prompt to picture | Anyone who can run Python | An image from a text description, from 512x512 up to 2048x2048 in total pixel area
Negative prompting | Artists tuning a result | Fewer unwanted artifacts, by naming what to leave out
Fine-tuning base | Researchers and community builders | A custom LoRA or ControlNet setup built on the base model

How to set it up:

• Python on Windows, macOS or Linux
• Room for about 12.3 GB of weights in BF16
• A recent diffusers release, which the model card asks for
• A CUDA GPU if you follow the card's example, which sends the model to cuda

1. Install the Hugging Face hub tools with pip.
2. Download the model files with the hf command.
3. Install the latest diffusers, as the model card describes.
4. Load the model in Python with ZImagePipeline.

pip install -U huggingface_hub
hf download Tongyi-MAI/Z-Image

How to use it: In Python you load the pipeline, give it a prompt and save the picture it returns. The card recommends a guidance scale of 3.0 to 5.0 and 28 to 50 inference steps.

import torch
from diffusers import ZImagePipeline

pipe = ZImagePipeline.from_pretrained(
    "Tongyi-MAI/Z-Image",
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=False,
)
pipe.to("cuda")

Tool | What it does
ZImagePipeline | Loads the Z-Image weights as a diffusers pipeline
prompt | Describes the picture you want
negative_prompt | Names what to keep out, and may stay empty
guidance scale | Recommended range is 3.0 to 5.0
inference steps | Recommended range is 28 to 50

Example: You ask for "two young women standing close together in front of a plain grey wall, both smiling at the camera, natural soft light, casual fashion photo". The model returns one image of that scene. The card's own example sets the height to 1280 and leaves the negative prompt empty.

For newcomers and experts: Start with the two download commands, then copy the short Python snippet and swap in a prompt of your own. Say plainly what you want to see, such as the place, the light and the mood. Leave the negative prompt empty on the first try.

Z-Image is a single-stream diffusion transformer that is not distilled, so it supports full classifier-free guidance and negative prompts. It takes 28 to 50 steps where Turbo takes 8, and the card rates its diversity high and its fine-tuning support yes. The trade is slower generation, and the card rates Turbo higher on visual quality.

Write a text prompt
Load ZImagePipeline in diffusers
Run 28 to 50 inference steps
Guidance scale steers the result
Get an image back

Good to know:

• License: Apache 2.0
• Cost to run: none named, but you need your own hardware and about 12.3 GB of weights in BF16
• Data it reaches: downloads the model files from Hugging Face, and the card's example then runs on your own machine
• Not for: fast eight-step generation, which the card gives to Z-Image-Turbo
• Limits: the card lists no reinforcement learning stage for this version

On Hugging Face, 13 November 2026

• Task: text to image
• Library: diffusers
• Parameters: 6.2B
• Weights: about 12.3 GB in BF16; about 3.1 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 45,594
• Likes: 1,302
• License: apache-2.0
• Languages: en
• Quantized versions: 28
• Fine-tunes: 72
• Tags: diffusers, safetensors, text-to-image

Links

• Model: https://huggingface.co/Tongyi-MAI/Z-Image
• Files: https://huggingface.co/Tongyi-MAI/Z-Image/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Tongyi-MAI/Z-Image

Then follow the usage in its model card. Nothing in this kit runs by itself.