Welcome to the official repository for the Z-Image（造相）project!
✨ Z-Image
Z-Image is a powerful and highly efficient image generation model family with 6B parameters. Currently there are four variants:
🚀 Z-Image-Turbo – A distilled version of Z-Image that matches or exceeds leading competitors with only 8 NFEs (Number of Function Evaluations). It offers ⚡️sub-second inference latency⚡️ on enterprise-grade H800 GPUs and fits comfortably within 16G VRAM consumer devices. It excels in photorealistic image generation, bilingual text rendering (English & Chinese), and robust instruction adherence.
🎨 Z-Image – The foundation model behind Z-Image-Turbo. Z-Image focuses on high-quality generation, rich aesthetics, strong diversity, and controllability, well-suited for creative generation, fine-tuning, and downstream development. It supports a wide range of artistic styles, effective negative prompting, and high diversity across identities, poses, compositions, and layouts.
🧱 Z-Image-Omni-Base – The versatile foundation model capable of both generation and editing tasks. By releasing this checkpoint, we aim to unlock the full potential for community-driven fine-tuning and custom development, providing the most "raw" and diverse starting point for the open-source community.
✍️ Z-Image-Edit – A variant fine-tuned on Z-Image specifically for image editing tasks. It supports creative image-to-image generation with impressive instruction-following capabilities, allowing for precise edits based on natural language prompts.
Showcase of Z-Image on Photo-realistic image Generation
📖 Accurate Bilingual Text Rendering: Z-Image-Turbo excels at accurately rendering complex Chinese and English text.
Showcase of Z-Image on Bilingual Text Rendering
💡 Prompt Enhancing & Reasoning: Prompt Enhancer empowers the model with reasoning capabilities, enabling it to transcend surface-level descriptions and tap into underlying world knowledge.
reasoning.jpg
🧠 Creative Image Editing: Z-Image-Edit shows a strong understanding of bilingual editing instructions, enabling imaginative and flexible image transformations.
Showcase of Z-Image-Edit on Image Editing
🏗️ Model Architecture
We adopt a Scalable Single-Stream DiT (S3-DiT) architecture. In this setup, text, visual semantic tokens, and image VAE tokens are concatenated at the sequence level to serve as a unified input stream, maximizing parameter efficiency compared to dual-stream approaches.
Architecture of Z-Image and Z-Image-Edit
📈 Performance
According to the Elo-based Human Preference Evaluation (on Alibaba AI Arena), Z-Image-Turbo shows highly competitive performance against other leading models, while achieving state-of-the-art results among open-source models.
Install the latest version of diffusers, use the following command:
Click here for details for why you need to install diffusers from source
We have submitted two pull requests (#12703 and #12715) to the 🤗 diffusers repository to add support for Z-Image. Both PRs have been merged into the latest official diffusers release.
Therefore, you need to install diffusers from source for the latest features and Z-Image support.
1import torch
2from diffusers import ZImagePipeline
34# 1. Load the pipeline5# Use bfloat16 for optimal performance on supported GPUs6pipe = ZImagePipeline.from_pretrained(7"Tongyi-MAI/Z-Image-Turbo",8 torch_dtype=torch.bfloat16,9 low_cpu_mem_usage=False,10)11pipe.to("cuda")1213# [Optional] Attention Backend14# Diffusers uses SDPA by default. Switch to Flash Attention for better efficiency if supported:15# pipe.transformer.set_attention_backend("flash") # Enable Flash-Attention-216# pipe.transformer.set_attention_backend("_flash_3") # Enable Flash-Attention-31718# [Optional] Model Compilation19# Compiling the DiT model accelerates inference, but the first run will take longer to compile.20# pipe.transformer.compile()2122# [Optional] CPU Offloading23# Enable CPU offloading for memory-constrained devices.24# pipe.enable_model_cpu_offload()2526prompt ="Young Chinese woman in red Hanfu, intricate embroidery. Impeccable makeup, red floral forehead pattern. Elaborate high bun, golden phoenix headdress, red flowers, beads. Holds round folding fan with lady, trees, bird. Neon lightning-bolt lamp (⚡️), bright yellow glow, above extended left palm. Soft-lit outdoor night background, silhouetted tiered pagoda (西安大雁塔), blurred colorful distant lights."2728# 2. Generate Image29image = pipe(30 prompt=prompt,31 height=1024,32 width=1024,33 num_inference_steps=9,# This actually results in 8 DiT forwards34 guidance_scale=0.0,# Guidance should be 0 for the Turbo models35 generator=torch.Generator("cuda").manual_seed(42),36).images[0]3738image.save("example.png")
🔬 Decoupled-DMD: The Acceleration Magic Behind Z-Image
Decoupled-DMD is the core few-step distillation algorithm that empowers the 8-step Z-Image model.
Our core insight in Decoupled-DMD is that the success of existing DMD (Distributaion Matching Distillation) methods is the result of two independent, collaborating mechanisms:
CFG Augmentation (CA): The primary engine 🚀 driving the distillation process, a factor largely overlooked in previous work.
Distribution Matching (DM): Acts more as a regularizer ⚖️, ensuring the stability and quality of the generated output.
By recognizing and decoupling these two mechanisms, we were able to study and optimize them in isolation. This ultimately motivated us to develop an improved distillation process that significantly enhances the performance of few-step generation.
Building upon the strong foundation of Decoupled-DMD, our 8-step Z-Image model has already demonstrated exceptional capabilities. To achieve further improvements in terms of semantic alignment, aesthetic quality, and structural coherence—while producing images with richer high-frequency details—we present DMDR.
Our core insight behind DMDR is that Reinforcement Learning (RL) and Distribution Matching Distillation (DMD) can be synergistically integrated during the post-training of few-step models. We demonstrate that:
If you find our work useful in your research, please consider citing:
bibtex
1@article{team2025zimage,
2 title={Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer},
3 author={Z-Image Team},
4 journal={arXiv preprint arXiv:2511.22699},
5 year={2025}
6}
78@article{liu2025decoupled,
9 title={Decoupled DMD: CFG Augmentation as the Spear, Distribution Matching as the Shield},
10 author={Dongyang Liu and Peng Gao and David Liu and Ruoyi Du and Zhen Li and Qilong Wu and Xin Jin and Sihan Cao and Shifeng Zhang and Hongsheng Li and Steven Hoi},
11 journal={arXiv preprint arXiv:2511.22677},
12 year={2025}
13}
1415@article{jiang2025distribution,
16 title={Distribution Matching Distillation Meets Reinforcement Learning},
17 author={Jiang, Dengyang and Liu, Dongyang and Wang, Zanyi and Wu, Qilong and Jin, Xin and Liu, David and Li, Zhen and Wang, Mengmeng and Gao, Peng and Yang, Harry},
18 journal={arXiv preprint arXiv:2511.13649},
19 year={2025}
20}
Quantized transformer checkpoints (this repo)
This repo adds pre-quantized diffusion transformer checkpoints for Tongyi-MAI/Z-Image-Turbo, built with torchao
dynamic activation quantization from the dense bf16 transformer. The official model card above
is unchanged from the source repo.
fp8: Float8DynamicActivationFloat8WeightConfig with PerRow granularity (e4m3, torch._scaled_mm).
The loader must floor the dynamic activation scale (activation_value_lb=1e-12 on torchao 0.13+)
so all-zero activation token rows cannot produce a zero scale.
Loading a checkpoint is bit-identical to quantizing the dense bf16 transformer on the fly;
the checkpoint skips the dense load and quantize step.
Validated against same-seed dense bf16 renders (SSIM, LPIPS-vgg, CLIP delta, non-finite and
black-frame checks) on torch 2.12.1 and torchao 0.17.
Samples
Prompt: "cute sloth typing on a computer" (1024x1024, family default steps/guidance, seeds 0-2).
int8
fp8
Pre-cast fp8 text encoder (this repo)
Z-Image-Turbo-text_encoder-FP8.pt (4.4 GB) is the pipeline's text encoder (Qwen3ForCausalLM,
Qwen3-4B, the text_encoder subfolder of Tongyi-MAI/Z-Image-Turbo) with the layerwise fp8
storage cast Unsloth Studio applies at load time, saved pre-cast:
Loading it is bit-identical to downloading the dense encoder and casting on load
(verified tensor for tensor: 399 tensors, 252 cast to fp8 storage; the tied
lm_head/embedding stays dense).
Cuts the text-encoder download from 8.0 GB to 4.4 GB.
Plain-tensor state dict: loads with torch.load(weights_only=True). Metadata records
scheme fp8, component text_encoder, base Tongyi-MAI/Z-Image-Turbo.
