← Back to Discover

Z-Image

Family of open 6B-parameter image generation models with a fast eight-step Turbo variant, a base model, an editing variant and an omni-capable foundation model.

Image generationOpen models
Visit website ↗

What it does

Z-Image is a family of image generation models with 6 billion parameters, released with inference code, checkpoints on Hugging Face and ModelScope, and an online demo. The repository lists four variants: Z-Image-Turbo, Z-Image, Z-Image-Omni-Base and Z-Image-Edit. Its model table marks the Omni-Base and Edit checkpoints as still to be released. Turbo is a distilled version that needs only eight function evaluations. The project says this gives sub-second inference on enterprise H800 GPUs and fits within 16 GB of VRAM on consumer devices, and it highlights photorealistic images and accurate English and Chinese text rendering. The base Z-Image model uses more steps and supports negative prompting and a wide range of styles, and is positioned for fine-tuning. You can run it with the inference scripts in the repository or through the Hugging Face diffusers library. The README also lists community tools that speed it up or cut memory use, including stable-diffusion.cpp, which it says can run Z-Image on as little as 4 GB of VRAM. The code is released under the Apache-2.0 license.

Capabilities

  • Z-Image-Turbo generates images in eight steps
  • Renders English and Chinese text in images
  • Base model with negative prompting and a wide style range
  • Editing variant for prompt-based image changes
  • Runs through PyTorch scripts or Hugging Face diffusers
  • Apache-2.0 license

When to use it

  • Generating images on your own GPU with a fast open model
  • Fine-tuning or building on an open image foundation model
  • Making images that contain Chinese or English text