Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer
Paper β’ 2511.22699 β’ Published β’ 245
import torch
from diffusers import DiffusionPipeline
# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("vantagewithai/Z-Image-GGUF", dtype=torch.bfloat16, device_map="cuda")
prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
image = pipe(prompt).images[0]Quantized GGUF version of Z-Image.
Original model link: https://huggingface.co/Tongyi-MAI/Z-Image
Watch us at Youtube: @VantageWithAI
β‘οΈ- Image
An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer
Welcome to the official repository for the Z-ImageοΌι ηΈοΌproject!
π¨ Z-Image
Z-Image is the foundation model of the β‘οΈ- Image family, engineered for good quality, robust generative diversity, broad stylistic coverage, and precise prompt adherence. While Z-Image-Turbo is built for speed, Z-Image is a full-capacity, undistilled transformer designed to be the backbone for creators, researchers, and developers who require the highest level of creative freedom.
π Key Features
- Undistilled Foundation: As a non-distilled base model, Z-Image preserves the complete training signal. It supports full Classifier-Free Guidance (CFG), providing the precision required for complex prompt engineering and professional workflows.
- Aesthetic Versatility: Z-Image masters a vast spectrum of visual languagesβfrom hyper-realistic photography and cinematic digital art to intricate anime and stylized illustrations. It is the ideal engine for scenarios requiring rich, multi-dimensional expression.
- Enhanced Output Diversity: Built for exploration, Z-Image delivers significantly higher variability in composition, facial identity, and lighting across different seeds, ensuring that multi-person scenes remain distinct and dynamic.
- Built for Development: The ideal starting point for the community. Its non-distilled nature makes it a good base for LoRA training, structural conditioning (ControlNet) and semantic conditioning.
- Robust Negative Control: Responds with high fidelity to negative prompting, allowing users to reliably suppress artifacts and adjust compositions.
π Z-Image vs Z-Image-Turbo
| Aspect | Z-Image | Z-Image-Turbo |
|---|---|---|
| CFG | β | β |
| Steps | 28~50 | 8 |
| Fintunablity | β | β |
| Negative Prompting | β | β |
| Diversity | High | Low |
| Visual Quality | High | Very High |
| RL | β | β |
Recommended Parameters
- Resolution: 512Γ512 to 2048Γ2048 (total pixel area, any aspect ratio)
- Guidance scale: 3.0 β 5.0
- Inference steps: 28 β 50
π Citation
If you find our work useful in your research, please consider citing:
@article{team2025zimage,
title={Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer},
author={Z-Image Team},
journal={arXiv preprint arXiv:2511.22699},
year={2025}
}
- Downloads last month
- 385
Hardware compatibility
Log In to add your hardware
2-bit
3-bit
4-bit
5-bit
6-bit
8-bit
16-bit