Z-Image Turbo

The first are the maker's own samples. The rest are this month's best-liked on Civitai, shown from there; each belongs to the person who made it. Open one for its maker.

About

Built from the Z-Image foundation model and trained to make a picture in eight steps. A prompt helper adds reasoning, so pictures can draw on world knowledge beyond the words given.

How it compares

In LM Arena's public vote, it places seventh of the 10 open picture models rated for making pictures.

Good to know

The project's Chinese name is written 造相.

Words and picture go in as one single stream instead of two separate ones.

Its researchers found a previously overlooked training part is the main reason few-step pictures work.

Sources: Tongyi-MAI/Z-Image-Turbo model card, Decoupled DMD: CFG Augmentation as the Spear, Distribution Matching as the Shield, Outmake's speed estimates, LM Arena and Hugging Face. Every line comes from one of these.

Run It Anywhere

Z-Image Turbo is an open model. Draw Things and DiffusionBee take it from Tongyi-MAI's own files on Hugging Face. For ComfyUI, with the ComfyUI-GGUF nodes installed, put one of the GGUF files in ComfyUI/models/unet.

• Draw ThingsMacOpen
• DiffusionBeeMacOpen
• ComfyUIWindows, Linux, MacFiles