Z-Image and Z-Image Turbo: Licence and How to Run Them

Z-Image is a family of open image generation models with 6 billion parameters, made by the Z-Image Team at Tongyi-MAI, part of Alibaba Group, and released under the Apache 2.0 licence. Z-Image-Turbo is its distilled version, which generates an image in 8 function evaluations, and you can run it natively in ComfyUI or with the Diffusers library.

Checked 13 October 2026 against the Z-Image repository and its licence file, the Z-Image-Turbo and Z-Image model cards, the technical report and the project page; ComfyUI's Z-Image-Turbo and Z-Image tutorials, its Z Image examples and Comfy Org's repackaged files; the Fun ControlNet Union card; and the model cards for FLUX.1 [schnell] and SDXL. An earlier version of this page quoted generation times and a VRAM range that no source supports. They have been removed, and we publish no speed benchmarks because we have not run any.

This page belongs to our AI image generation for influencers workflow. For how Z-Image sits beside the other families you can load, see best ComfyUI models.

What is Z-Image?

Z-Image is described by its authors as "a powerful and highly efficient image generation model family with 6B parameters". The technical report positions it against larger open models: it says leading alternatives have 20 to 80 billion parameters, which makes them impractical on consumer hardware, and presents Z-Image as a 6 billion parameter model built on a Scalable Single-Stream Diffusion Transformer. In that design, text tokens and image tokens are joined into one sequence and processed together.

The repository lists four variants. Two are downloadable today.

Read the status column before following an older tutorial. The project page talks about Z-Image-Edit, and ComfyUI's docs list it as a variant, but the model table in the repository and on the model card still marks both Z-Image-Edit and Z-Image-Omni-Base as "To be released". There is no official editing checkpoint to download yet.

Who makes Z-Image, and what is the licence?

The project page credits "Z-Image Team, Tongyi MAI, Alibaba Group". The weights are published by the Tongyi-MAI organisation on Hugging Face and on ModelScope.

The licence file in the repository is the Apache License, Version 2.0, and both model cards carry the Apache 2.0 tag. That is one of the most permissive licences among current image models: our SDXL guide sets the families and their licences side by side. Apache 2.0 covers the weights and code. It does not decide what you may publish: likeness law and platform rules still apply to the images.

Z-Image-Turbo vs Z-Image: which one to run

For persona work the split is practical. Turbo is the lane for fast drafts: composition, outfit and lighting tests where you want many images quickly. Its low diversity rating means seeds resemble each other more, which helps when you want a steady look and hurts when you want options. The base model is the lane for anything you need to steer with a negative prompt, and for training.

How to run Z-Image-Turbo in ComfyUI

ComfyUI supports Z-Image natively: its README lists the model, and the official tutorial ships a workflow in the template library.

Sizes are from Comfy Org's Hugging Face repositories. The same repository lists smaller versions of the Turbo diffusion model (6.2 GB and 4.51 GB files) and of the text encoder (5.63 GB and 3.48 GB files). No custom nodes are needed for the basic workflow. If you add community packs later, install them through ComfyUI Manager.

Pose and depth control exist too. ComfyUI's tutorial includes a workflow for the Z-Image-Turbo Fun Union ControlNet, which supports Canny, HED, Depth, Pose and MLSD and goes in ComfyUI/models/model_patches. Our ControlNet guide explains the control types, and the img2img guide covers refining a draft in a second pass.

How to run Z-Image-Turbo with Diffusers

The repository's quick start installs Diffusers from source, then loads the model with the ZImagePipeline class:

pip install git+https://github.com/huggingface/diffusers

import torch
from diffusers import ZImagePipeline

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

image = pipe(
    prompt=prompt,
    height=1024,
    width=1024,
    num_inference_steps=9,  # This actually results in 8 DiT forwards
    guidance_scale=0.0,     # Guidance should be 0 for the Turbo models
    generator=torch.Generator("cuda").manual_seed(42),
).images[0]

The comments are the README's own, and prompt is your text. For memory-constrained devices the same example offers pipe.enable_model_cpu_offload(). For the base model, load Tongyi-MAI/Z-Image and use the recommended parameters: 28 to 50 steps, a guidance scale of 3.0 to 5.0, and a resolution from 512x512 to 2048x2048.

Z-Image hardware requirements

The team makes one statement about consumer hardware: Z-Image-Turbo "fits comfortably within 16G VRAM consumer devices". It also reports sub-second inference on an enterprise H800 GPU, which says little about a home graphics card. ComfyUI's README says it can run even the biggest open source models on as low as 4GB of VRAM with 8GB of RAM, a stated minimum and not a speed promise; the project's other documented requirements are on our ComfyUI statistics page.

Z-Image-Turbo vs FLUX.1 schnell vs SDXL

These are each publisher's own statements. We have not timed or scored the models against each other.

Training a persona LoRA on Z-Image

Turbo will not give you the same face twice from a prompt alone; no base model does. For a persona you add a character LoRA, and the model cards point to the base model for that: Z-Image's "non-distilled nature makes it a good base for LoRA training". AI Toolkit, OneTrainer and musubi-tuner list Z-Image, and our LoRA training guide covers the dataset and settings. The free LoRA training steps calculator works out the length of the run.

The subject of that LoRA is a fictional adult persona that you created, or yourself. An Apache 2.0 licence on the model does not permit training on, or reproducing, a real person's face without their written permission. The AI Influencers program covers the full persona pipeline on current models, from the trained face to video and the disclosure rules for publishing.

Z-Image FAQ