Chinese company Z.ai announces open-source image generation AI 'GLM-Image,' a hybrid of autoregressive and diffusion models



Z.ai , a Chinese AI company, has announced ' GLM-Image ,' an open-source discrete autoregressive image generation model that delivers industrial-grade performance. GLM-Image is a hybrid of an autoregressive model and a diffusion model .

GLM-Image: Auto-regressive for Dense-knowledge and High-fidelity Image Generation
https://z.ai/blog/glm-image



Z.ai announced their image generation model, GLM-Image, on January 14, 2026 (local time). It is a hybrid model that combines an autoregressive model and a diffusion model. The autoregressive model is initialized based on GLM-4-9B-041 , which has 9 billion parameters. The diffusion model follows CogView4 and employs a single-stream DiT structure with 7 billion parameters.

Hybrid GLM-Image models offer significant advantages in text rendering and knowledge-intensive generation, particularly in tasks requiring precise semantic understanding and complex information representation, while enabling generation that is faithful to the prompt.

Below is a summary of an example of image generation using GLM-Image.



Diffusion models have become the mainstream image generation model due to their training stability and strong generalization ability. Although diffusion models and

variational autoencoders (VAEs) have been improved year by year, diffusion models still struggle with complex instructions and knowledge-intensive scenarios.

On the other hand, in recent years, there has been an increase in image generation AI that uses autoregressive models, which enable output that is faithful to prompts. However, autoregressive models have the disadvantage of slow execution. GLM-Image was developed as a hybrid model that combines the advantages of diffusion models and autoregressive models.

In GLM-Image, an autoregressive generator produces tokens carrying low-frequency semantic signals, while a diffusion decoder refines high-frequency details to provide the final image. This hybrid architecture not only works reliably for general image generation tasks, but also offers significant advantages in creative tasks requiring complex knowledge representation.

The diagram below shows the image generation mechanism of GLM-Image.



Below is a table comparing the text rendering accuracy of image generation models. GLM-Image is an open source model, yet it achieves overwhelmingly high text rendering accuracy.

Model Open Source NED CLIPScore Word Accuracy 2 regions Word Accuracy 3 regions Word Accuracy 4 regions Word Accuracy 5 regions Word Accuracy average
GLM-Image 〇 0.9557 0.7877 0.9103 0.9209 0.9169 0.8975 0.9116
Seedream 4.5 × 0.9483 0.8069 0.8778 0.8952 0.9083 0.9008 0.899
Z-Image 〇 0.9367 0.7969 0.9006 0.8722 0.8652 0.8512 0.8671
Qwen-Image-2512 〇 0.929 0.7819 0.863 0.8571 0.861 0.8618 0.8604
Z-Image-Turbo 〇 0.9281 0.8048 0.8872 0.8662 0.8628 0.8347 0.8585
GPT Image 1[High] × 0.9478 0.7982 0.8779 0.8659 0.8731 0.8218 0.8569
Seedream 4.0 × 0.9224 0.7975 0.8585 0.8484 0.8538 0.8269 0.8451
Qwen-Image 〇 0.9116 0.8017 0.837 0.8364 0.8313 0.8158 0.8288
Nano Banana 2.0 × 0.8754 0.7372 0.7368 0.7748 0.7863 0.7926 0.7788
TextCrafter 〇 0.8679 0.7868 0.7628 0.7628 0.7406 0.6977 0.737
SD3.5 Large 〇 0.847 0.7797 0.7293 0.6825 0.6574 0.594 0.6548
Seedream 3.0 × 0.8537 0.7821 0.6282 0.5962 0.6043 0.561 0.5924
FLUX.1 [dev] 〇 0.6879 0.7401 0.6089 0.5531 0.4661 0.4316 0.4965
3DIS 〇 0.6505 0.7767 0.4495 0.3959 0.388 0.3303 0.3813
RAG-Diffusion 〇 0.4498 0.7797 0.4388 0.3316 0.2116 0.191 0.2648
TextDiffuser-2 〇 0.4353 0.6765 0.5322 0.3255 0.1787 0.0809 0.2326
AnyText 〇 0.4675 0.7432 0.0513 0.1739 0.1948 0.2249 0.1804


Since GLM-Image is an open-source image generation model, the code and model data are publicly available on GitHub and Hugging Face.

GitHub - zai-org/GLM-Image: GLM-Image: Auto-regressive for Dense-knowledge and High-fidelity Image Generation.
https://github.com/zai-org/GLM-Image



zai-org/GLM-Image · Hugging Face
https://huggingface.co/zai-org/GLM-Image



Related Posts:

in AI, Posted by logu_ii