The dark horse of AI image generation, reshaping the tech landscape with "small but beautiful". When other AIs need high-end servers, your RTX 3060 can smoothly generate high-quality images.
Enter a prompt to start generating
Z-image是于2025年11月发布的高效图像生成基础模型，核心定位是"轻量且高性能"。个人和企业可自由使用、修改并二次分发，搭配官方提供的技术报告和快速入门代码，方便开发者进行二次开发。
高效推理版（已发布）
基础开发版（待发布）
图像编辑版（待发布）
Single-stream diffusion Transformer, unified processing of text, vision, and image VAE tokens, breaking through traditional dual-stream architecture bottlenecks
Unified rotary position encoding, perfectly adapted to multi-dimensional position information of text and images
Ensures stable convergence of thousand-layer networks, solves deep network gradient problems
Compresses inference steps to 8, combines reinforcement learning to patch details, achieves both speed and quality
RTX 4090: Generate 1024×1024 in just 2.3 seconds
RTX 3060: Graphics cards from 5 years ago can still run smoothly
Compared to Flux 2, saves 80% storage space
FID Score: 7.2 (lower is better)
AI Arena Ranking: Global 4th
CVTG-2K Accuracy: 0.8671
Precisely understands artistic conceptions like "small bridge, flowing water, homes"
Handles complex Chinese-English mixed instructions
Overcomes text blur and character error issues
Chinese tea house wooden sign with clear "Ming Xiang" characters, antique architecture, red lanterns, bamboo decorations, morning sunlight
Young Asian woman at cafe entrance, wearing beige sweater, gentle smile, glass door reflection, afternoon sunlight, 85mm lens
Cyberpunk style city night scene, neon billboards, rain-soaked street reflections, Eastern elements, futuristic
Z-image excels at processing Chinese-English mixed prompts for more precise descriptions
Specify "clear text", "font" for better results
1024×1024, 8-9 steps, Guidance 0.0
Consumer-grade RTX 3060 graphics cards can run smoothly. RTX 4090 generates 1024×1024 images in just 2.3 seconds with 13GB memory usage.
If you encounter insufficient memory:
Individuals, studios and enterprises can use it for free commercially. When distributing after secondary development, keep copyright notices and follow relevant license regulations.
If generated images are used commercially, conduct content review, avoid adult content, infringing elements and other violations; also suggest keeping the prompts and parameter records used for generation to handle possible copyright checks.