Tongyi-MAI /
Z-Image-Turbo
A 6B-parameter text-to-image model that makes photorealistic pictures in 8 steps.
- 5.5k likes
- 697.0k downloads a month
- 6.2B parameters
- Some experience
- Image generation
Z-Image-Turbo is the fast version of the Z-Image family, which has 6B parameters. It is a distilled model, meaning it was trained to get good results in only 8 function evaluations instead of the 50 steps the base Z-Image uses. The card says it runs in under a second on enterprise H800 GPUs and fits within 16G of VRAM on consumer devices. It uses a single-stream diffusion transformer, where text and image tokens go through the model as one sequence.
The card lists some limits in its own table. Turbo has low diversity, so different prompts or seeds may give less varied results than the base Z-Image. Guidance is not used (CFG is off), and fine-tunability is listed as N/A, so if you want to fine-tune, the card points to the base Z-Image instead. Z-Image-Edit and Z-Image-Omni-Base are listed as 'To be released', so Turbo is for generating images, not editing them. You also need to install diffusers from source.
What you need to run it
6.2B parameters. The weights take about 12 GB at full 16-bit precision (fits a 16 GB Mac or a GPU with 16 GB), or about 4 GB as a 4-bit download, which runs on a 8 GB Mac or a GPU with 8 GB. Long inputs and images need more memory on top.
What it does well
Few steps, fast output
It needs only 8 NFEs, with sub-second latency on H800 GPUs according to the card.
Fits on consumer GPUs
The card says it fits comfortably within 16G VRAM consumer devices, and the example code mentions optional CPU offloading for memory-constrained machines.
Photorealism
The card highlights strong photorealistic image generation with good aesthetic quality.
Bilingual text in images
It is described as accurate at rendering complex Chinese and English text.
Follows instructions
The card lists robust instruction adherence, and says it ranks as state-of-the-art among open-source models in an Elo-based human preference evaluation on Alibaba AI Arena.
What you could build
- Generate photorealistic pictures from text prompts for a class project or portfolio.
- Make posters or mock-ups that include English or Chinese text.
- Build a small image generator app on top of the diffusers library.
- Try different prompts and seeds on a 16G VRAM GPU to learn how text-to-image prompting works.
- Use it as a quick image generator in a prototype where waiting for 50 steps would be too slow.
Get started
Install diffusers from source, then load ZImagePipeline with the Tongyi-MAI/Z-Image-Turbo checkpoint and generate an image using num_inference_steps=9 and guidance_scale=0.0, as in the card's Python example. To download the weights ahead of time, install huggingface_hub and run the download command.
pip install git+https://github.com/huggingface/diffuserspip install -U huggingface_hubHF_XET_HIGH_PERFORMANCE=1 hf download Tongyi-MAI/Z-Image-TurboCommands are copied from the model card. Read it on Hugging Face for the full setup.