GLM-Image is an open-source discrete auto-regressive image generation model that combines a 9B parameter semantic token generator with a 7B DiT diffusion decoder for high-fidelity, knowledge-dense image creation and editing.
What is GLM-Image?
GLM-Image is a hybrid image generation model that accepts text prompts or reference images and outputs high-resolution visuals (1024×1024 to 2048×2048 pixels). It runs on the Z.ai platform and is developed by Z.AI. The model uses an auto-regressive generator to encode semantic structure and a diffusion decoder to restore fine details, enabling accurate text rendering and complex layouts.
Key Features
- Hybrid AR + Diffusion Architecture — A 9B-parameter auto-regressive transformer produces semantic tokens, while a 7B DiT (Diffusion Transformer) decoder reconstructs textures, lighting, and fine-grained realism.
- Semantic-VQ Tokenization — Preserves layout and meaning in the token sequence, improving convergence and enabling knowledge-dense compositions.
- Accurate Text Rendering — Glyph-byT5 tokenizer supports readable text, including Chinese characters, in generated images.
- Text-to-Image and Image-to-Image — Supports generation from prompts, editing existing images, style transfer, identity preservation, and multi-subject consistency.
- High-Resolution Output — Trained on resolutions from 1024×1024 to 2048×2048 pixels, suitable for print and production workflows.
- Open-Source Access — Model weights, training code, and inference tools are available on GitHub and pre-trained checkpoints on Hugging Face under a license permitting research and commercial use.
Who is it for?
- Content creators — Generate posters, signage, or social media visuals with readable typography and accurate layouts.
- Researchers and developers — Fine-tune or deploy the open-source checkpoints for custom image-generation pipelines.
- Design professionals — Use image-to-image editing and style transfer to iterate on high-fidelity outputs for client projects.
What can you do with GLM-Image?
- Create knowledge-dense visuals — Input a complex prompt with multiple objects, text, and spatial relationships; get a coherent, high-resolution image with accurate text and layout.
- Edit and transform existing images — Upload a reference image for style transfer, identity preservation, or multi-subject consistency across generated variations.
- Generate production-ready graphics — Export results at up to 2048×2048 pixels for use in print, web, or digital content without immediate post-processing.
How does GLM-Image work?
- Define the goal — Describe the subject, constraints, and any text the image must contain.
- Choose generation mode — Select text-to-image for creation or image-to-image for edits and variations.
- Generate and iterate — Refine prompts to improve text rendering, layout, and details.
- Export high-resolution results — Save outputs at 1024–2048 pixels via the web interface or API.
Pricing
GLM-Image is freemium. New users receive 15 free credits (💎) after logging in. Additional credits can be purchased through the pricing page. The model weights and code are freely available for self-hosted use under an open-source license.
FAQ
Is GLM-Image free to use online?
Yes, you can try GLM-Image on the Z.ai platform with 15 free credits after login. Each generation consumes one credit. For heavier usage, paid credit packs are available.
What resolutions does GLM-Image support?
The model supports output resolutions from 1024×1024 to 2048×2048 pixels. The web interface offers a preset of 1280×1280 (Square 1:1), and other aspect ratios can be selected.
