Z-Image Turbo
A practical 6B image generation model
The Z-Image project explores how far a compact model can go in image generation when the focus is on design and training efficiency instead of size alone. Z-Image Turbo is the main distilled branch of this effort.
Why Z-Image was built
Many image generation systems rely on very large models. They produce strong results but are out of reach for most researchers, students, and small teams. Z-Image takes a different path: it aims to show that well chosen architectures, balanced data, and careful training can make a 6‑billion‑parameter model competitive in everyday tasks.
Z-Image Turbo is the distilled variant of this work. It is tuned to generate images with only a short sampling schedule while still preserving structure and detail. The project treats runtime cost as an important factor, so that running local experiments on a single GPU remains realistic.
Alongside Turbo, Z-Image Edit extends the same backbone for controlled editing tasks. Together, these two branches cover a wide range of common workflows: creating new images from prompts and adjusting existing ones with clear instructions.
What this project provides
Z-Image Turbo is more than a single checkpoint. It is a set of tools and resources that make it easier to understand, run, and adapt the model.
Model weights
The weights for Z-Image Turbo and Z-Image Edit are released so that you can inspect how the models behave, test them on your own prompts, and integrate them into custom pipelines.
Reference code
The reference implementation shows how to load checkpoints, run text encoders, and carry out the sampling loop. It is written with clarity in mind so that it can also serve as learning material.
Online demo
For people who want to try the model before installing anything locally, a simple demo is available. It helps users understand prompt behavior, bilingual text rendering, and sampling settings.
Design principles
Efficiency first
The model is sized so that it runs on widely available hardware. This allows more people to explore it in detail, run their own tests, and keep it in their regular workflow.
Clarity
The project aims to keep the training and inference logic understandable. Clear code and simple examples make it easier for new users to follow how images are produced.
Accessibility
By releasing both code and weights together with an online demo, the project lowers the barrier for people who want to learn from or extend modern diffusion models.
How to use this site
Homepage
The homepage gives a high‑level overview of Z-Image Turbo, highlights the main ideas, and points you to the demo and installation notes.
Installation
The installation page provides step‑by‑step guidance on setting up an environment, loading checkpoints, and generating your first images locally.
Blog
The blog section shares short notes, examples, and observations about prompts, bilingual rendering, and experiments with Z-Image Turbo and Z-Image Edit.
Project direction
The Z-Image project will continue to focus on models that balance capability with practical deployment. Future updates may add new sampling methods, more guidance options, or additional editing tools, but the aim will remain the same: keep the models understandable, efficient, and suitable for a wide range of everyday setups.
Feedback from users is an important part of this work. Real‑world prompts, usage patterns, and observations help reveal where the model works well and where it needs to improve. This site will evolve to reflect those lessons over time.