Z-Image
Scalable, single-stream DiT architecture-based image generation model.
- S3-DiT Architecture: Scalable Single-Stream Diffusion Transformer — Integrates text, visual semantic tokens, and image VAE latent variables into a single sequence, maximizing parameter efficiency compared to Dual-Stream. Achieves quality comparable to 20B+ commercial models with 6B parameters. - Multi-Resolution Generation: Supports arbitrary aspect ratios from 512×512 to 2048×2048, with native 1024×1024 as the default. - 8-Step Fast Generation (Turbo): Completes generation in just 8 NFE (Number of Function Evaluations) using the Decoupled-DMD distillation algorithm (CFG Augmentation + Distribution Matching). Achieves inference in under 1 second on enterprise GPUs and is compatible with consumer 16GB VRAM devices. - Dual-Language Text Rendering: Accurately renders Chinese and English text within images, achieving text consistency comparable to closed-source commercial models. - Image-to-LoRA (i2L): Automatically extracts LoRA weights from reference images, enabling style transfer. - CFG Precise Control: The base model fully supports negative prompts, allowing for fine-grained adjustment of the realism-to-stylization spectrum within the guidance scale range of 3.0–5.0. - Image Editing (Z-Image-Edit): Instruction-following editing based on Omni pre-training, supporting dual-language editing commands.
💻 System Requirements
Minimum 16GB (based on bf16 inference). Can be reduced to 4GB with stable-diffusion.cpp quantization. With enterprise GPUs (H800/A100 level), inference time is less than 1 second. Low-spec systems can be supported with the CPU offloading option.
Approximately 12GB for the model checkpoint (based on bf16), approximately 20GB including the text encoder (Qwen-3 4B) and VAE.
⚡ Installation
### 4-1. Quick Start
```bash
# Install the latest diffusers source (Z-Image pipeline support required)
pip install git+https://github.com/huggingface/diffusers
# Download the model (using high-speed XET transfer)
HF_XET_HIGH_PERFORMANCE=1 huggingface-cli download Tongyi-MAI/Z-Image
```
### 4-2. Detailed Installation (PyTorch Native)
```bash
git clone https://github.com/Tongyi-MAI/Z-Image.git
cd Z-Image
pip install -e .
python inference.py
```
### 4-3. Turbo Model Inference (High-Speed, 8 Steps)
```python
import torch
from diffusers import ZImagePipeline
pipe = ZImagePipeline.from_pretrained(
"Tongyi-MAI/Z-Image-Turbo",
torch_dtype=torch.bfloat16
).to("cuda")
image = pipe(
prompt="A photorealistic portrait of a scientist in a lab",
height=1024, width=1024,
num_inference_steps=9,
guidance_scale=0.0
).images[0]
image.save("output.png")
```
### 4-4. Base Model Inference (CFG Supported)
```python
import torch
from diffusers import ZImagePipeline
pipe = ZImagePipeline.from_pretrained(
"Tongyi-MAI/Z-Image",
torch_dtype=torch.bfloat16
).to("cuda")
image = pipe(
prompt="Fluorescence microscopy image of neural cells",
height=1024, width=1024,
num_inference_steps=30,
guidance_scale=4.0
).images[0]
```🧬 Bio Use Cases
Large-Scale Generation of Synthetic Training Data
Generate a large number of synthetic microscopy images of rare cell types (e.g., circulating tumor cells) based on prompts, to address the data imbalance problem in deep learning models for cell classification. The CFG (guidance scale 3.0–5.0) of the base model can be controlled to balance morphological diversity and realism. LoRA fine-tuning enables the mass production of images specialized in specific staining patterns (H&E, DAPI, etc.).
Automated Generation of Bilingual Scientific Posters and Infographics
Leveraging its ability to render Chinese and English text, this tool generates draft versions of conference presentation posters and graphical abstracts for research papers with a single prompt. Rendering at a high resolution of 2048x2048 without text distortion minimizes the need for post-processing. Particularly useful for creating bilingual posters for international conferences.
ControlNet-Based Experimental Environment Simulation
By combining ControlNet Union 2.1's Depth/Canny/Pose modes, this tool generates simulation images that maintain the composition of existing laboratory photos while modifying lighting, background, and equipment layout. It can be used for laboratory design reviews, safety training material creation, and equipment catalog visuals. Licensed under Apache-2.0, allowing for commercial use without restrictions.
📝 Update Notes
No update notes yet.
🧪 Related Code of Life
No related Code of Life posts yet.