Z-Image
Z-Image is a family of open-weight text-to-image models from Alibaba's Tongyi Lab, built on a compact ~6B architecture that punches well above its size on prompt adherence, clean composition, and legible in-image text — including Chinese and English. The Turbo variant renders in just a handful of steps, while the Base model trades speed for maximum fidelity. Generate with both right here on Civitai — no GPU, no install.
About Z-Image
Z-Image is an open-weight text-to-image family from Alibaba's Tongyi Lab (released under the Tongyi-MAI banner), built on a compact ~6-billion-parameter architecture. Despite its small size it targets photorealistic image generation, bilingual text rendering in both English and Chinese, and robust instruction adherence — the kind of prompt-following that usually demands a much larger model. Because the weights are lightweight, the Turbo variant fits comfortably within 16GB of consumer VRAM and reaches sub-second latency on enterprise H800 GPUs.
The family ships as three checkpoints tuned for different jobs. Z-Image Turbo is a distilled model that produces a finished image in just 8 sampling steps (NFEs), making it the fast, few-step default; Z-Image Base is the non-distilled foundation model, released to unlock the full quality ceiling and to give the community a clean base for fine-tuning and custom development; and Z-Image Edit is a separate variant tuned for instruction-driven image-to-image editing. On Civitai, Turbo and Base are both hosted, so you can iterate quickly on Turbo and switch to Base when you want maximum fidelity — no downloads, no local GPU.
Reach for Z-Image when you want faithful prompt adherence and clean, legible in-image text — especially mixed English/Chinese text — at a fraction of the cost and wait of heavier models. Turbo is the natural pick for fast drafting and high-volume work, while Base rewards patience with the full step count for its best output. For the deepest style and character LoRA libraries, the SDXL-based Pony and Illustrious ecosystems still lead; Z-Image's own fine-tune library is smaller but growing, and its speed-to-quality ratio makes it a strong everyday text-to-image workhorse.
How to prompt Z-Image
- Write in natural language and follow Z-Image's 6-part structure: Subject, Scene, Composition, Lighting, Style, Constraints — in that order. Lead with the subject (and any text you want rendered), since it matters most.
- Keep prompts short: attention fades after roughly 75 tokens (about 50–60 words), so front-load the important content and trim trailing detail that will otherwise be ignored.
- Do not use negative prompts. Turbo is a few-step distilled model with no CFG at inference, so every constraint has to be phrased positively inside the main prompt — describe what you want, not what you want to avoid.
- Skip weight syntax like (word:1.3) — it is not supported. Control emphasis through word order and description instead.
- For photorealism, lighting is the single strongest lever — be specific about it — and add sensory detail such as "skin texture," "fabric detail," "imperfections," or "film grain." For text in the image, Z-Image renders English and Chinese directly, so just state the words you want.
Featured Z-Image models
Curated — the models worth generating with first.
Popular Z-Image LoRAs & add-ons
Top LoRAs by downloads — live data, refreshed daily. Stack them on any checkpoint.
Example generations
Curated, safe-for-work showcase — every image ships with its prompt and settings.
How to run Z-Image
Two paths — one takes ten seconds, one takes an afternoon.
⚡ Run on Civitai (Recommended)
The fastest way to start.
🖥️ Run it locally
For power users who want full control.
No graphics card? The Civitai path above skips all of this.
Z-Image vs other ecosystems
Backed by Civitai usage data.
| Feature | Z-Image | Flux | SDXL | Qwen |
|---|---|---|---|---|
| Best for | Fast, faithful text-to-image | Photorealism & text | General purpose, speed | Prompt adherence & text |
| Prompt adherence | Very good | Excellent | Good | Excellent |
| Text in images | Strong (EN + CN) | Strong | Weak | Strong |
| Speed on Civitai | Fastest (few-step) | Fast (4–8s) | Fast (2–4s) | Medium (6–12s) |
| LoRA ecosystem | 9,000+ | 40,000+ | 38,000+ | Growing |
| Available on Civitai | ✓ Yes | ✓ Yes | ✓ Yes | ✓ Yes |
Frequently asked questions
How much does it cost to generate with Z-Image?
What's the difference between Z-Image Turbo and Z-Image Base?
Who made Z-Image?
Can I train my own Z-Image LoRA?
Do I need a GPU to run Z-Image?
Start generating with Z-Image now
No installation. No GPU. Runs in your browser.