dee-z-image

dee-z-image

This repository hosts a text-to-image checkpoint in Diffusers format. It is compatible with ZImagePipeline and can be loaded directly from the Hugging Face Hub.

Usage

Install

Install the latest Diffusers (recommended) and the required runtime dependencies:

pip install -U torch transformers accelerate safetensors
pip install -U diffusers

If your installed Diffusers version does not include ZImagePipeline, install Diffusers from source instead:

pip install -U git+https://github.com/huggingface/diffusers

Generate an image

import torch
from diffusers import ZImagePipeline

model_id = "telcom/dee-z-image"

pipe = ZImagePipeline.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,  # use torch.float16 if your GPU does not support bf16
    low_cpu_mem_usage=False,
)
pipe.to("cuda")

prompt = "A cinematic studio photo of a small robot sitting at a desk, warm lighting, shallow depth of field, high detail."

image = pipe(
    prompt=prompt,
    height=1024,
    width=1024,
    num_inference_steps=9,
    guidance_scale=0.0,
    generator=torch.Generator("cuda").manual_seed(42),
).images[0]

image.save("out.png")

Tips

• If you run out of VRAM, try pipe.enable_model_cpu_offload() (requires accelerate) or reduce the resolution.
• Start with guidance_scale=0.0 and num_inference_steps around 8–12; adjust based on quality/speed needs.
• For reproducibility, set a generator seed as shown above.

Repository contents

• model_index.json defines the Diffusers pipeline components used by ZImagePipeline.
• text_encoder/, tokenizer/, transformer/, vae/, scheduler/ contain the model submodules.
• assets/ contains example images and an optional gallery PDF.

License

Apache-2.0 (see metadata at the top of this model card).

Acknowledgements

This repo packages a checkpoint for the Z-Image family of models. For upstream project details, see:

• https://github.com/Tongyi-MAI/Z-Image
• https://arxiv.org/abs/2511.22699