Z-Image

Z-Image is a lightweight image generation tool featuring an efficient 8-Step inference architecture. It delivers fast, high-quality AI image generation on consumer-grade GPUs while significantly reducing computational costs.

1 credit = 1 image 10 welcome credits Private by default

Create, remix, and refine in one place

Start simple, then open precise controls only when you need them.

0/1500

Style preset

Preview a look. Presets are free; only generation uses credits.

Dimensions

1 credit = 1 standard image Downloads are always free
Sign in to claim 10 free credits.

What are you creating today?

Purpose-built starting points for common creative goals.

Chinese text poster

Accurate bilingual headlines, signage, and campaign layouts.

Product photography

Clean e-commerce sets and premium studio campaigns.

Editorial portrait

Art-directed portraits with camera, light, and composition control.

Social campaign

Platform-ready vertical creative with clear type hierarchy.

A result is the start, not the finish

Every saved image carries its prompt, dimensions, seed, credit cost, and parent reference into the next step.

01

Generate

Build the first direction

02

Remix

Use it as a reference

03

Refine

Control how much changes

04

Reframe

Adapt the composition

05

Download

Export without charge

Use what is ready now. See what comes next.

See credit pricing

Text to image

Ready

Simple and advanced generation with visual style presets.

1 credit

Image remix

Ready

Enhance, refine, restyle, or reimagine a private source.

1 credit

Prompt builder

Ready

Add realism, camera, lighting, composition, and exact text.

Free

HD enhance

Ready

Continue from a result and preserve its lineage.

1 credit

Remove background

Next

Transparent PNG output for products and assets.

1 credit

Inpaint canvas

Ready

Brush a mask and replace only the selected region.

4 credits

Start creating

ControlNet

Next

Guide pose, depth, or edges from a reference.

3 credits

Not Just Fast, It's Fully Evolved

Filling the gap between lightweight and massive models, Z-Image-Turbo finds the perfect balance between speed, quality, and usability.

Native Bilingual Support

Powered by Qwen 3.4B LLM. No more garbled Chinese characters. Calligraphy, signage, and complex typography are rendered precisely.

S3-DiT Single Stream

Radical architectural innovation. Text and image tokens are processed consistently, similar to GPT-4, utilizing every parameter for both generation and understanding.

Apache 2.0 License

True open-source freedom. Unlike Flux.1's commercial restrictions, you are free to use it commercially, modify, and integrate. Ideal for startups and game studios.

6B Parameters Golden Balance
8 Steps Inference Decoupled-DMD
Qwen 3.4B Text Encoder Native Bilingual
12GB VRAM Req No Quantization Needed

S3-DiT: Breaking Modal Barriers

Traditional models use a "dual-stream" architecture. Z-Image-Turbo adopts Scalable Single-Stream Diffusion Transformer (S3-DiT).

  • Unified Input Stream: Text Tokens and Image Latents are concatenated directly.
  • Full Parameter Interaction: Every Transformer layer performs deep text-image attention calculation.
  • Decoupled-DMD: The core algorithm that compresses inference to just 8 steps.
  • CFG Enhancement: Independently optimized guidance signals for sharp images without high CFG values.
Architecture_v1.0
Text Token
+
Img Latent
Unified Transformer Block Self-Attention (All-to-All)
High-Fidelity Output (8 Steps)

Why Choose Z-Image-Turbo?

We provide the optimal solution balancing performance, cost, and ecosystem.

MetricZ-Image-TurboFlux.1 (Dev)SDXL Base
Parameters6B (Balanced)12B (Massive)2.6B
VRAM Req12GB (Native BF16)24GB+ (or Quant)8GB
Steps8 Steps (Distilled)20-50 Steps20-50 Steps
Text EncoderQwen 3.4B (Bilingual)T5 + CLIPOpenCLIP
Typography⭐️⭐️⭐️⭐️⭐️ Perfect⭐️⭐️ Poor⭐️ Garbled
LicenseApache 2.0Non-CommercialOpenRAIL++
Cost/ImgFrom $0.005 (API)HighLow

A Boon for Consumer Hardware

Thanks to the 6B parameter scale and 8-step distillation, Z-Image-Turbo achieves 2-3s generation on RTX 3090/4090. For enterprise H800s, sub-second response is reality.

Nvidia H800 (Enterprise) < 1 s
RTX 4090 (Consumer High-End) ~ 2.5 s
Flux.1 Dev (RTX 4090) ~ 10 s+

Quick Start

# Quick load with Diffusers

from diffusers import DiffusionPipeline

import torch


# Load 8-Step Turbo Model

pipe = DiffusionPipeline.from_pretrained(

  "Tongyi-MAI/Z-Image-Turbo",

  torch_dtype=torch.bfloat16

).to("cuda")


# Generate Image

image = pipe(

  prompt="Cyberpunk detective, rainy night, neon lights, Chinese sign saying "Tongyi Lab"",

  num_inference_steps=8,

  guidance_scale=1.0 # Distilled models don't need high CFG

).images[0]

Frequently Asked Questions

Questions about deployment, usage, and licensing.

GPU requirements?

For native precision (BF16), 16GB VRAM (RTX 4080/3090) is recommended. With GGUF/NF4 quantization, 8GB VRAM cards (RTX 3060) run smoothly with minimal quality loss.

Can I use it commercially?

Yes. Z-Image-Turbo uses the permissive Apache 2.0 license. You can use it freely for commercial products without fees.

How to write Chinese prompts?

Just like chatting naturally. Thanks to Qwen 3.4B, you can use complex sentences, idioms, or poems. For text rendering, wrap specific text in quotes.

Support for ComfyUI / WebUI?

Yes. ComfyUI has Day-0 support (update to latest). Automatic1111 support is in the dev branch and coming soon.

Advantage over Flux.1?

Z-Image-Turbo solves efficiency and usability. While Flux is great for extreme quality, Z-Image offers 3x speed, half VRAM usage, and superior Chinese support.