Chinese text-to-image models

China's generative landscape is dominated by high-efficiency models like Z-Image-Turbo and CogView4 that outperform Western counterparts in bilingual text rendering and inference speed.

Chinese text-to-image technology has shifted from rapid following to global leadership by prioritizing architectural efficiency and native bilingual support. Key players like Alibaba Cloud and Zhipu AI have released models such as Z-Image-Turbo (6B parameters) and CogView4 (6B parameters) that leverage Diffusion Transformer (DiT) architectures to achieve sub-second inference on enterprise hardware. These systems excel in rendering complex Chinese characters and traditional aesthetic styles (shanshui) where models like Stable Diffusion often struggle. Recent benchmarks place Alibaba's Z-Image-Turbo as the top open-source model on the Artificial Analysis leaderboard (December 2025), while Baidu's ERNIE-ViLG 2.0 utilizes a 24-billion parameter mixture-of-denoising-experts to maintain a 77% win rate in human preference evaluations for high-fidelity scene composition.

https://github.com/Tongyi-MAI/Z-Image