Guides

Z-Image Turbo Guide: What Z-Image Is and How to Run It

Z-Image is Alibaba Tongyi's 6B open image model. What Z-Image Turbo is, its licence, VRAM needs, official settings and how to prompt it.

By Anna Krejcova — Editor at FansCreatorPublished 11 min read
Article summary
  • Z-Image is a 6-billion-parameter open image model family from Alibaba's Tongyi team. Z-Image Turbo came out on 26 November 2025 and the base Z-Image on 27 January 2026, both under Apache 2.0.
  • Z-Image Turbo makes a picture in 8 steps, fits on consumer cards with 16 GB of VRAM, and runs with guidance at 0, so negative prompts do nothing on it.
  • Official Turbo settings: 9 steps in Diffusers (8 model passes), guidance 0.0, 1024x1024. The base model wants 28 to 50 steps, guidance 3.0 to 5.0 and a negative prompt.
  • Prompt it with long, concrete descriptions in English or Chinese, put on-image text in double quotes, and skip tags like 8K or masterpiece.

Z-Image is a family of open-weight image models from Alibaba’s Tongyi Lab, published under the Tongyi-MAI name. Its best-known version, Z-Image Turbo, makes a picture in 8 steps, fits on consumer graphics cards with 16 GB of video memory, and is released under the Apache 2.0 licence. You can download it from GitHub, Hugging Face or ModelScope, try it in an official online demo, or call it through Alibaba Cloud’s API.

Every number below comes from the official GitHub repository, the Hugging Face model cards, the technical report or Alibaba Cloud’s documentation, all read on 5 October 2026.

What is Z-Image?

Z-Image (造相 in Chinese) is a text-to-image model with 6 billion parameters. You write a description, and it returns a picture. The project site credits the “Z-Image Team, Tongyi MAI, Alibaba Group”. On ModelScope, Alibaba’s model hub, the Tongyi-MAI organization goes by the Chinese name 通义多模态交互, Tongyi Multimodal Interaction.

The point of Z-Image is size. Its technical report notes that leading open models such as Qwen-Image, Hunyuan-Image-3.0 and FLUX.2 have 20 to 80 billion parameters, which makes them “impractical for inference, and fine-tuning on consumer-grade hardware”. Z-Image aims for similar quality at 6 billion. The report says the full training run took 314,000 H800 GPU hours, about 630,000 US dollars.

It uses what the team calls a Scalable Single-Stream Diffusion Transformer (S3-DiT): text and image data enter the model as one stream instead of two. The text encoder in the Hugging Face repository is a Qwen3 language model.

Z-Image release timeline

Date What happened
26 November 2025 Z-Image Turbo released on Hugging Face and ModelScope
27 November 2025 First version of the technical report posted to arXiv
8 December 2025 Z-Image Turbo ranked 8th overall and first among open-source models on the Artificial Analysis text-to-image leaderboard, per the README
27 January 2026 Z-Image, the full foundation model, released

Z-Image model variants

The GitHub README lists four variants. Two can be downloaded; the other two were still marked “To be released” as of 5 October 2026.

Variant What it is for Steps Guidance (CFG) Status on 5 Oct 2026
Z-Image Turbo Fast, photorealistic text-to-image 8 No Released
Z-Image Foundation model, styles, fine-tuning 28 to 50 Yes Released
Z-Image-Omni-Base Raw base for generation and editing 50 Yes To be released
Z-Image-Edit Editing images with text instructions 50 Yes To be released

Turbo is a distilled version of Z-Image. Distillation teaches a smaller or faster model to copy a slower one, so it needs fewer steps. The README says Turbo “matches or exceeds leading competitors with only 8 NFEs” (8 passes through the model).

Z-Image vs Z-Image Turbo

The Z-Image model card compares the two released models directly:

Z-Image Z-Image Turbo
Steps 28 to 50 8
Guidance (CFG) Yes No
Negative prompts Yes No
Fine-tuning Yes No
Diversity across seeds High Low
Visual quality High Very high

The short version: Turbo is for fast, polished pictures. The base model is for control, variety and training your own add-ons. The official card marks Turbo as not meant for fine-tuning and calls the base model “a good base for LoRA training”, so train LoRAs on the base model. Community finetunes of Turbo exist anyway: the model tree on the Turbo Hugging Face page listed over 150 finetunes and over 800 adapters as of 5 October 2026. They are not official releases.

Z-Image licence

Both Hugging Face model cards list the licence as Apache 2.0, and the GitHub repository ships the Apache 2.0 licence file. Apache 2.0 is a permissive licence. It allows commercial use, changes and redistribution, as long as you keep the licence and its notices with what you share.

The licence covers the code and the weights, not the people in your pictures or the rules of the site where you post.

Hardware: what you need to run Z-Image Turbo

The official numbers are short:

  • Video memory: the README says Turbo “fits comfortably within 16G VRAM consumer devices”. The technical report says “less than 16GB VRAM”.
  • Speed: “sub-second inference latency” on an H800, which is a data-centre GPU. A home card will be slower.
  • Precision: the example code loads the model in bfloat16.
  • Low memory: the example code includes an optional CPU offloading line “for memory-constrained devices”.
  • Disk space: the weight files in the Turbo repository add up to about 33 GB.

The README’s community list says stable-diffusion.cpp, a C++ inference engine, can run Z-Image “with as little as 4GB of VRAM”. That is a community tool, not the official release.

How to run Z-Image Turbo

1. Try the online demo

The fastest way to see Z-Image is the official demo on Hugging Face Spaces. ModelScope hosts one too. No install needed.

In the Hugging Face demo, steps are locked at 8 and the “Time Shift” slider defaults to 3.0. Resolution presets run from 1024x1024 up to 2048x1152, in ratios from 1:1 to 21:9. The demo’s code checks each prompt for unsafe content and runs a safety checker on the result, so it is a general-audience tool.

2. Run it in Python with Diffusers

This is the route the Turbo model card documents in full. The example code runs on a CUDA graphics card.

  1. Install Diffusers from source: pip install git+https://github.com/huggingface/diffusers. The README explains that Z-Image support arrived through two pull requests, so it asks for the latest code.
  2. Download the weights: pip install -U huggingface_hub, then hf download Tongyi-MAI/Z-Image-Turbo.
  3. Load the ZImagePipeline in bfloat16 and move it to the GPU.
  4. Generate with the official settings.

The core of the official example looks like this:

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
).images[0]

The comments are the team’s own.

3. Run it in ComfyUI

ComfyUI, the open-source node-based interface, is a common community route. The README’s community list includes a ComfyUI node that sets up the official Z-Image resolutions. ComfyUI’s own examples page lists three files for Z-Image Turbo: the qwen_3_4b text encoder, the z_image_turbo_bf16 diffusion model and the Flux 1 VAE. You put each in its ComfyUI models folder, then load the example workflow. Use the same settings as below: 8 steps and guidance off.

4. Run the native PyTorch code

Clone the GitHub repository, install it with pip install -e ., then run python inference.py. This uses the team’s own code instead of Diffusers.

5. Call it through Alibaba Cloud

Alibaba Cloud Model Studio offers z-image-turbo as a hosted API model. Its model page lists two regions, China (Beijing) and Singapore, and a limit of 120 requests per minute, as of 5 October 2026. It suits apps without their own GPU.

These are the settings the official model cards give.

Setting Z-Image Turbo Z-Image (base)
Steps 9 in Diffusers (8 model passes) 28 to 50
Guidance scale 0.0 3.0 to 5.0
Negative prompt Not used “Strongly recommended”
Resolution 1024x1024 in the example; demo presets up to 2048x1152 512x512 to 2048x2048 total pixel area, any aspect ratio
CFG normalization Not used False for stylised work, True for realism
Scheduler shift (repository config, 5 Oct 2026) 3.0 6.0

Guidance at 0 matters for prompting. Guidance is what makes a negative prompt work, so on Turbo a negative prompt has nothing to act on. Put everything you want into the main prompt instead.

How to prompt Z-Image

Z-Image reads long, plain descriptions in English or Chinese. The official examples are full paragraphs, not tag lists. The best guide to what it likes is the team’s own prompt enhancer, a template in the demo that rewrites short prompts into full ones. Translated from Chinese, it works in this order:

  1. Lock the core. Subject, number of people or objects, action, state, colours and any named text stay exactly as asked.
  2. Add the photo details. Composition, light and mood, materials and texture, colour scheme, and layers of depth.
  3. Quote any text. Words that should appear in the picture go in English double quotes, written exactly. For posters and menus, describe the font and the layout too.
  4. Stay concrete. No metaphors, no emotional language, and no meta tags such as “8K” or “masterpiece”.

A prompt built that way reads like a description of a real photo:

Photo of a woman in her thirties reading at a café window, soft morning light from the left, white ceramic cup on a wooden table, shallow depth of field, street blurred behind the glass, a chalkboard sign on the wall that says “OPEN 8-6”, muted warm colours.

Three more tips follow from the model cards:

  • Use text. The README highlights “bilingual text rendering (English & Chinese)”. Signs, labels and posters are a strength, so name the words in quotes.
  • Vary the prompt, not just the seed. The base card rates Turbo’s diversity as low. A new seed tends to give a similar picture, so change the wording when you want a different composition.
  • Describe, don’t forbid. With no negative prompt on Turbo, write “natural skin texture with visible pores” rather than listing what you don’t want.

For a full method of camera, lens and light terms, see prompt engineering for realistic AI photos.

Strengths and limits of Z-Image Turbo

Strengths, as the official sources describe them:

  • Photorealism. The README says Turbo “excels in photorealistic image generation”.
  • Text in images, in English and Chinese.
  • “Robust instruction adherence”, in the README’s words.

Limits to plan around:

  • No negative prompts and no guidance. You steer Turbo only through the main prompt.
  • Low diversity. The model card’s own rating. Good for consistency of style, less good for exploring ideas.
  • Not meant for fine-tuning. The official card marks Turbo that way; train LoRAs on the base Z-Image model.
  • No official editing model yet. Z-Image-Edit and Z-Image-Omni-Base were still listed as “To be released” as of 5 October 2026.
  • No memory of a person. Like any text-to-image model, each generation starts fresh, so the same character will drift from picture to picture unless you add a consistency method.
  • A general-audience release. The technical report says the training data was scored for NSFW content as part of filtering, and the official demo screens prompts.

Z-Image for creators who post as one person

The last two limits matter most if you sell content, because subscribers follow one face. Methods for holding a face steady, from LoRA training to reference images, are compared in LoRA training vs multi-angle prompting.

FansCreator lists Z-Image among the models it runs, and it keeps one creator’s face and body the same across images and video.

For adult creators, FansCreator’s NSFW AI image generator and Z-Image page are built for adult content with consent rules. You download the result and post it yourself.

Whatever tool you use, the consent rule is the same: only your own face, your own AI persona, or a face whose owner gave consent you can prove. The FansCreator Terms of Use forbid anything else. Then check where the picture is going:

  • OnlyFans allows AI content only if it “clearly features the verified Creator” and is labelled, for example with #ai (OnlyFans Help Centre). Its moderation guide lists content that is “wholly AI-generated” or “features heavily altered images” among what it removes. A persona built from scratch does not belong there.
  • Fanvue allows fully AI-generated creators with a clear AI disclosure in the bio, the caption or a watermark (Fanvue Help Centre).
  • MYM allows AI content labelled “Virtual Image”, #AI or #AIgenerated, and adds an “AI profile” badge (MYM support).
  • Fansly does not allow photorealistic AI content, even when it is labelled (Fansly Help Centre).

If you are building an original persona for Fanvue or MYM, how to create an AI influencer walks through the steps.

Turbo or base model: which Z-Image to pick

Use Z-Image Turbo when you want realistic pictures fast: 8 steps, guidance at 0, about 1024x1024 and a long, concrete prompt. Use the base Z-Image model when you need negative prompts, more varied compositions or a base for LoRA training, and accept 28 to 50 steps. Both are Apache 2.0, and neither keeps a person consistent on its own.

Frequently asked questions

Who makes Z-Image?

Z-Image comes from Alibaba. The project site credits the Z-Image Team, Tongyi MAI, Alibaba Group, and the models are published under the Tongyi-MAI name on GitHub, Hugging Face and ModelScope.

What is the difference between Z-Image and Z-Image Turbo?

Z-Image Turbo is a distilled version of Z-Image built for speed: 8 steps, no guidance, no negative prompts and very high visual quality, but low variety between seeds. The base Z-Image needs 28 to 50 steps, supports guidance and negative prompts, gives more varied results and is the one to fine-tune.

How much VRAM does Z-Image Turbo need?

The official README says Z-Image Turbo fits comfortably within 16 GB of VRAM on consumer devices, and the technical report says less than 16 GB. The example code also offers CPU offloading for cards with less memory, at the cost of speed.

Can I use Z-Image commercially?

Yes. Both Z-Image and Z-Image Turbo are released under the Apache 2.0 licence, which allows commercial use, modification and redistribution as long as you keep the licence and notices. The licence does not override the rights of people shown in your images or the rules of the platform where you post.

What are the best settings for Z-Image Turbo?

The model card uses 9 inference steps in Diffusers, which equals 8 passes through the model, a guidance scale of 0.0, bfloat16 precision and 1024x1024 resolution. Change the prompt rather than the guidance to steer the result.

Sources

Read as Markdown

Bring your next idea to life.

Create images, videos and content for your audience with FansCreator.

Start for Free