GLM Image

GLM-Image is an AI-powered image generation model developed by zai-org that adopts a hybrid autoregressive and diffusion decoder architecture. Positioned in line with mainstream latent diffusion approaches in general image generation quality, the tool offers notable benefits in scenarios necessitating text-rendering and knowledge-intensive generation. It demonstrates impressive performance in tasks that require robust semantic understanding and intricate information expression, while ensuring high-fidelity and detailed image generation. The architecture involves a 9B-parameter autoregressive generator, initializing from GLM-4-9B-0414 with additional visual tokens, a diffusion decoder, and a post-training system with the reinforcement learning algorithm GRPO to augment both semantic understanding and visual detail quality. GLM-Image is equipped to handle both text-to-image and image-to-image generation. It offers capabilities to generate high-detail images from textual descriptions, and supports a wide array of image-to-image tasks including image editing, style transfer, consistent generation of multiple subjects, and identity-preserving generation.