The base model of the image generation AI 'Z-Image' has finally arrived. It's 'strong at illustrations,' 'produces a variety of faces and compositions,' and 'is ideal for additional learning.'



Tongyi-MAI, Alibaba's AI development team, released the image generation AI ' Z-Image ' on January 28, 2026. Z-Image is the base model of

Z-Image-Turbo , which was released in November 2025, and it is expected that a variety of models will be produced through fine tuning. Compared to Z-Image-Turbo, Z-Image is also characterized by its ability to output illustration-style images with higher quality and a greater variety of compositions and characters.




The Z-Image series is an image generation AI series consisting of 'Z-Image,' 'Z-Image-Turbo,' 'Z-Image-Omni-Base,' and 'Z-Image-Edit.' Of these, only Z-Image-Turbo was released in November 2025.

Alibaba releases 'Z-Image,' a high-speed, high-quality image generation AI - GIGAZINE



Z-Image-Turbo is a model developed using the procedure of 'fine-tuning Z-Image-Omni-Base to create a Z-Image, distilling the Z-Image, and then applying reinforcement learning with human feedback.' While it has the characteristic of 'high-speed and high-quality image generation,' it also has the disadvantage of 'low diversity in output images, making it unsuitable for additional learning and creating unique models.' The Z-Image released this time is a pre-distillation model, characterized by high diversity and ease of fine-tuning.



Below is a comparison table of Z-Image and Z-Image-Turbo. Z-Image boasts features such as 'easy fine tuning,' 'negative prompt support,' and 'high versatility.' However, the number of steps required for the generation process increases from 28 to 50, and the overall image quality drops one level from 'Very High' to 'High.'



Below are some examples of Z-Image. While the specs suggest that the quality is lower than Z-Image-Turbo, it still produces high-quality images.



The quality has improved significantly, especially for illustrative images.



Z-Image-Turbo had problems with generating the same composition even when the seed value was changed, and generating images containing multiple people with the same faces, but Z-Image makes it possible to generate diverse images.



It also supports negative prompts, allowing you to explicitly specify elements you do not want to include in the image.



The Z-Image model data is available at the following link:

Tongyi-MAI/Z-Image · Hugging Face
https://huggingface.co/Tongyi-MAI/Z-Image



ComfyUI also already supports image generation using Z-Image.




Related Posts:

in AI, Posted by log1o_hf