Z-Image Turbo

Key facts

What sets it apart

• Speed: the Turbo variant is distilled and needs only eight steps. On data-centre GPUs an image takes less than a second.
• Photorealism: skin texture, light and materials look far less polished than in many other models.
• Text in the image: English and Chinese lettering is rendered reliably and legibly – signs, posters, even tattoos.
• Modest requirements: officially 16 GB of VRAM is enough; quantised GGUF versions go well below that.

The price of speed: results vary little from seed to seed. To get variety, change the prompt rather than just the random value.

Running it locally with stable-diffusion.cpp

Instead of ComfyUI, Z-Image Turbo can run on stable-diffusion.cpp, a lean C++ implementation. Its Vulkan backend also works on AMD graphics cards without ROCm. Three files are needed:

• Diffusion model: e.g. z_image_turbo-Q5_K_M.gguf (Q4_0 approx. 3.7 GB, Q8_0 approx. 6.6 GB)
• Text encoder: Qwen3-4B-Instruct-2507 in GGUF format
• VAE: ae.safetensors from FLUX.1

Recommended settings: 8 steps, CFG 1.0, Euler sampler, flash attention enabled for the diffusion model. On a card with only 4 GB of VRAM (e.g. a Radeon RX 6400) this works if the weights stay in system memory and are streamed to the GPU in portions:

sd-server.exe ^
  --diffusion-model z_image_turbo-Q5_K_M.gguf ^
  --llm Qwen3-4B-Instruct-2507-Q4_K_M.gguf ^
  --vae ae.safetensors ^
  --backend "diffusion=vulkan0,te=cpu,vae=cpu" ^
  --params-backend "diffusion=cpu,te=disk,vae=cpu" ^
  --flash-attn off,diffusion=on ^
  --vae-tiling ^
  --max-vram vulkan0=2.0 ^
  --width 768 --height 768 ^
  --steps 8 --cfg-scale 1

Note: LoRAs made for Flux or Krea do not work. Z-Image needs LoRAs trained specifically for it, often labelled “ZIT” on platforms such as Civitai.

Prompting

Z-Image Turbo uses a large language model as its text encoder, so it understands detailed, naturally phrased descriptions better than keyword lists. A fixed structure works well:

1. Subject: who or what, with concrete details.
2. Setting: place, background, secondary figures.
3. Light and camera: flash, daylight, lens, grain.
4. Style and era: e.g. “early 2010s party snapshot” or “1990s documentary photography”.

Negative prompts have practically no effect at CFG 1. Rather than describing what should not be in the image, describe precisely what should. Put lettering in quotation marks.

Open source – or just open?

The weights and the inference code are released under the Apache 2.0 licence. The model may be used, modified, redistributed and used commercially – unlike Flux Dev, for instance, whose licence rules out commercial use. The training data and the data pipeline, however, are not published. Strictly speaking it is an “open weights” model – which is true of almost every major image model.

The question of training data

In the technical report the team describes an elaborate data pipeline that profiles images, removes duplicates and filters out low-quality material. The developers stress that they curated real data rather than using synthetic images from other models. Where the images actually come from is not disclosed.

The community draws conclusions mainly from the model’s behaviour: it knows celebrities and well-known imagery and filters far less than commercial services. This points to a broad, rather loosely filtered web dataset. Whether it contains the work of particular photographers or archives cannot be verified from the outside. For anyone who works with images, that remains the open ethical question behind the impressive technology.

Sources and links

• arXiv 2511.22699
• GitHub: Tongyi-MAI/Z-Image
• Hugging Face: Z-Image-Turbo
• Online demo
• stable-diffusion.cpp