✦
Summary
GLM-Image is an advanced image generation model distinguished by its innovative hybrid autoregressive and diffusion decoder architecture. This dual approach allows it to achieve visual quality comparable to mainstream latent diffusion models while exhibiting substantial superiority in tasks demanding dense knowledge and precise text rendering. Its architecture is carefully constructed, featuring a large autoregressive generator initialized from a powerful language model foundation, expanded to handle visual tokens, which initially crafts a compact encoding that is then elaborated into high-resolution image details.
Key Features
Hybrid autoregressive + diffusion decoder architecture for combined strengths.
9B-parameter autoregressive generator with expanded visual token vocabulary.
7B-parameter diffusion decoder equipped with a specialized Glyph Encoder for superior text rendering.
Post-training with GRPO for decoupled reinforcement learning, enhancing semantics and detail.
Strong performance in text-rendering and knowledge-intensive generation scenarios.
Supports high-fidelity, fine-grained detail generation.
Comprehensive image-to-image capabilities including editing and style transfer.
Supports identity-preserving generation and multi-subject consistency.
About GLM Image | z.AI
The sophistication of GLM-Image is further enhanced by its unique post-training regimen utilizing decoupled reinforcement learning, specifically the GRPO algorithm. This modular feedback strategy tailors learning signals based on the component they influence: the autoregressive module receives low-frequency feedback to refine aesthetic quality and semantic adherence to the prompt, ensuring better instruction following. Concurrently, the diffusion decoder receives high-frequency feedback specifically to improve detail fidelity and the accuracy of rendered text within the image, leading to exceptionally realistic textures and sharp textual elements.
Beyond standard text-to-image generation, GLM-Image is designed for versatility, seamlessly supporting a comprehensive suite of image-to-image capabilities within the same framework. This encompasses nuanced image editing, complex style transfer operations, and critical tasks like identity-preserving generation for maintaining consistent appearances of subjects across multiple outputs. The model’s ability to excel in information-intensive scenarios, such as creating complex layouts with accurate embedded text, positions it as a powerful tool for applications requiring both high visual fidelity and semantic precision.