GLM-Image

GLM-Image utilizes a unique model architecture that combines a 9B-parameter autoregressive generator, initialized from GLM-4-9B-0414 and expanded with visual tokens, with a 7B-parameter diffusion decoder based on a single-stream DiT architecture for latent-space image decoding.

The diffusion decoder is equipped with a Glyph Encoder text module, which significantly improves the accuracy and quality of text rendering within generated images, making it suitable for information-dense visual content.

The model incorporates a post-training phase using a fine-grained, modular feedback strategy based on the GRPO algorithm. This approach substantially enhances both semantic understanding and visual detail quality, with separate feedback for aesthetics/semantic alignment (autoregressive module) and detail fidelity/text accuracy (decoder module).

GLM-Image can generate high-detail images from textual descriptions, demonstrating particularly strong performance in scenarios that involve information-dense or knowledge-intensive prompts.

The model supports a wide array of image-to-image tasks within a single framework, including image editing, style transfer, identity-preserving generation for people and objects, and maintaining multi-subject consistency.

GLM-Image is fully integrated with the Hugging Face `transformers` and `diffusers` libraries, allowing for easy installation and usage via standard Python pipelines for both text-to-image and image-to-image tasks.