What is Z-Image?

Z-Image is a family of image generation models by Alibaba / Tongyi-MAI.

Since the name Z-Image itself refers to the whole model family, it's a little confusing, but this page covers Z-Image as the base model the others are derived from (sometimes called Z-Image-Base to tell them apart).

As a base model (the starting point for fine-tuning), Z-Image behaves in a plain, unbiased way.

It doesn't have the stabilizing distillation and reinforcement learning that Z-Image-Turbo has, so differences in seed and initial noise show up readily in the output. That gives it more creativity and variety, but it's also a tricky model: it's sensitive to parameters and results can swing a lot.


Model Download

📂ComfyUI/
└── 📂models/
    ├── 📂diffusion_models/
    │   └── z_image_bf16.safetensors
    ├── 📂text_encoders/
    │   └── qwen_3_4b.safetensors
    └── 📂vae/
        └── ae.safetensors

text2image

z_image.json
  • steps : It depends on the sampler, but a slightly higher count, around 30-40, is more stable.

Refine with Z-Image-Turbo

This refines Z-Image's output with Z-Image-Turbo in a few steps. The goal is to get both Z-Image's creativity and Z-Image-Turbo's consistent quality.

You could just do image2image, but here let's be a little fancy and split the sampling into two stages.

z_image_refine_turbo.json

Here we split it into the first 50% and the last 50%. (cf. Split Sampling)

  • 🟪 Z-Image : 15 steps out of 30 steps
  • 🟨 Z-Image-Turbo : 4 steps out of 8 steps
Outputs
Z-Image only
Z-Image only
Z-Image + Turbo
Z-Image + Turbo

Z-Image-Fun-Controlnet-Union-2.1

A ControlNet-like patch for Z-Image.

Model Download

📂ComfyUI/
└── 📂models/
    └── 📂model_patches/
        └── Z-Image-Fun-Controlnet-Union-2.1.safetensors

workflow

z_image_fun_controlnet_union_2_1.json
  • 🟩 Add the model and control image to QwenImageDiffsynthControlnet.
  • 🟩 In this workflow, Depth Anything V2 is used to create a depth map.

Reference