Z-Image

Z-Image packs competitive text-to-image quality into a 6B-parameter single-stream diffusion transformer from Alibaba's Tongyi MAI team, whose S3-DiT design concatenates text, visual semantic, and image VAE tokens into one input sequence. The distilled Turbo variant generates in 8 function evaluations without classifier-free guidance, reaching sub-second latency on an H800 and fitting consumer GPUs with 16 GB of VRAM, while the undistilled foundation model runs at 28 to 50 steps with guidance scales of 3.0 to 5.0. Z-Image-Omni-Base covers combined generation and editing, and Z-Image-Edit handles instruction-following image edits. A signature strength is accurate rendering of complex Chinese and English text inside images, which the team reports as comparable to top-tier closed-source models. Everything is Apache-2.0, weights are on Hugging Face and ModelScope with online demos, and a technical report is on arXiv. The authors claim photorealism on par with models an order of magnitude larger, and the repository has climbed to roughly 11.8k stars.