GLM-Image

Visit Website

GLM-Image (z.ai/blog/glm-image, released by Z.ai / Zhipu AI) is an open-source, 16B-parameter hybrid visual generation model. By combining a 9B discrete autoregressive generator (initialized from GLM-4-9B) for global layout planning and semantic understanding with a 7B diffusion decoder for high-frequency detail refinement, GLM-Image sets a benchmark for accurate multilingual text rendering (English & Chinese), complex scientific/knowledge-dense diagrams, and multi-subject image editing at resolutions up to 2048px.

 

Pure diffusion architectures frequently struggle with complex prompt logic, dense layout arrangement, and legible multi-line text embedding, while pure autoregressive image models often fall short on fine texture fidelity and skin rendering. GLM-Image resolves this trade-off with a decoupled two-stage pipeline: the autoregressive core generates low-frequency semantic visual tokens (semantic-VQ), and the diffusion decoder synthesizes high-frequency visual textures.

Technical Architecture & Model Matrix

Feature / Dimension GLM-Image Specification Primary Capability Advantage
Total Architecture Size 16 Billion Parameters (9B AR + 7B Diffusion) Strong semantic comprehension combined with fine visual generation.
Autoregressive Core GLM-4-9B foundation + Semantic-VQ tokenization Handles global scene planning, token hierarchy, and composition layout.
Diffusion Decoder 7B High-Fidelity Diffusion Refinement Network Recovers missing high-frequency details, edges, and photorealistic textures.
Text Rendering Accuracy 91.16% on CVTG-2k (EN) / 97.88% on LongText-Bench (ZH) Industry-leading text legibility across multi-line posters and infographics.
Generation Tasks Text-to-Image, Image Editing, Style Transfer, Identity Preservation Multi-turn editing conditioning on semantic tokens and reference VAE latents.
Output Resolution 1024px up to 2048px (aspect-ratio flexible) Native high-resolution output without external super-resolution artifacts.
Open-Source Availability Open Weights (Hugging Face, GitHub, vLLM / SGLang) Free for local research, community fine-tuning, and enterprise deployment.

Key Capabilities and Technical Innovations

  • Decoupled Autoregressive + Diffusion Pipeline: The 9B autoregressive transformer acts as a visual “planner,” outputting compact discrete tokens that establish composition, object placement, and typography logic. The 7B diffusion model then decodes and denoises these tokens into final high-resolution images.

  • State-of-the-Art In-Image Typography: Specifically excels at rendering legible words, paragraphs, and slogans inside quotation marks—making it ideal for marketing posters, scientific figures, user interface mockups, and Chinese/English bilingual designs.

  • Disentangled Reinforcement Learning (RL) Post-Training: Employs separate reward modeling for each stage: low-frequency rewards (HPSv3 aesthetics and VLM alignment) guide the autoregressive planner, while high-frequency perceptual rewards (LPIPS detail scoring and hand-correctness models) guide the diffusion decoder.

  • Knowledge-Dense & Educational Visualizations: Generates accurate technical illustrations, labeled anatomical diagrams, flowcharts, and academic presentations that adhere strictly to multi-step domain prompts.

VISIT WEBSITE