Z-Image (Zao Xiang) is an open-source image generation foundation model family from Alibaba Tongyi Lab, with 6B parameters, adopting a scalable single-stream DiT (S3-DiT) architecture that concatenates text, visual semantic tokens, and image VAE tokens at the sequence level into a unified input stream to improve parameter efficiency.
Uses and variants: Z-Image-Turbo is a distilled version, requiring only 8 function evaluations (NFE), achieving sub-second inference on enterprise-grade H800 GPUs, and can run on consumer-grade devices with 16G VRAM. It excels at realistic images, Chinese-English bilingual text rendering, and instruction following; Z-Image is the base model, focusing on high quality, rich aesthetics, diversity, and controllability, suitable for creative generation, fine-tuning, and downstream development, supporting multiple artistic styles, negative prompts, and high diversity; Z-Image-Omni-Base supports both generation and editing, aimed at community fine-tuning and custom development; Z-Image-Edit is fine-tuned for image editing, supporting precise image-to-image editing based on natural language prompts.
Main capabilities: realistic image quality, accurate Chinese-English bilingual text rendering, prompt enhancement and reasoning, creative image editing. The official claim is that Z-Image-Turbo ranks 8th on the Artificial Analysis text-to-image leaderboard, first among open-source models, and leads in Alibaba AI Arena human preference evaluations.
Usage: The model weights have been released on Hugging Face and ModelScope, with an online Demo also provided. The official repository provides instructions for virtual environment and dependency installation, and for running generation code; PRs (#12703, #12715) have also been submitted to diffusers and merged, requiring installation of diffusers from source to get Z-Image support. Recommended parameters: resolution 512×512 to 2048×2048 (by total pixel area, any aspect ratio), guidance scale 3.0–5.0, inference steps 28–50, strongly recommend using negative prompts, CFG normalization generally False for stylization, True for realism.
Community ecosystem: Cache-DiT, stable-diffusion.cpp (can run on as little as 4GB VRAM), LeMiCa, ComfyUI ZImageLatent, DiffSynth-Studio, vllm-omni, SGLang-Diffusion, Candle, MeanCache, etc. provide acceleration, training, and deployment support.
Limitations: Specific license, commercial terms, pricing, and complete installation commands are not yet provided, and the official repository should be taken as authoritative.
