Skip to content

Z-Image

Model Source Revision Weight license Parity (CPU, fp32)
z-image Tongyi-MAI/Z-Image 04cc4ab Apache-2.0 Verified
z-image-turbo Tongyi-MAI/Z-Image-Turbo f332072 Apache-2.0 Verified

Catalog: z-image · Models: z-image, z-image-turbo · Task: t2i

Architecture. A Qwen3-4B text encoder renders the prompt through its chat template and supplies penultimate hidden states. A 6B single-stream DiT uses 3-axis RoPE and refiner blocks. The FLUX-style VAE has 16 latent channels and a shift factor. z-image is the non-distilled base model with the same architecture. The flow-match Euler scheduler uses shift 6.0 for z-image and shift 3 for z-image-turbo. The native image size is 1024×1024.

Usage.

python -m imagehub.pretrained z-image-turbo checkpoints/z-image-turbo
from imagehub import pipeline

with pipeline("t2i", backend="z-image", model="checkpoints/z-image-turbo", device="mps") as pipe:
    result = pipe("a cat, oil painting", seed=42)
    result.images[0].save("cat.png")

Settings. For z-image, steps defaults to 50 and guidance defaults to 4.0. For z-image-turbo, they default to 9 and 0.0. width/height default to 1024 for both and must be multiples of 16. The supported options are negative_prompt, num_images, seed, latents, cfg_normalization, cfg_truncation and max_sequence_length. Unrecognized settings raise an error.

Seed compatibility. On CPU, the same seed gives the same image as diffusers with generator=torch.Generator("cpu").manual_seed(seed) (validation). MPS and CUDA runs are smoke tests only; there is no claim of exact equality on these devices.

Out of scope. There is no safety checker or invisible watermark. Omni, edit, controlnet, inpaint, img2img and fp16/bf16 parity are not in this release.

License. The weights are licensed under Apache-2.0. imagehub does not distribute weights.

Reference and imagehub, side by side

Drag the slider. Left is the recorded reference, right is imagehub, from the same seed, on CPU in float32.