🚀

Image Generation - GLM-Image

🛠

Text-to-Image & Image-to-Image Generation

Generate high-quality images from text prompts or existing images. GLM-Image is a hybrid autoregressive + diffusion model with exceptional text-rendering accuracy, strong semantic understanding, and support for both text-to-image and a wide range of image-to-image tasks

What it does

GLM-Image is an image generation model built on a hybrid autoregressive + diffusion decoder architecture. It matches mainstream latent diffusion approaches in general image quality while showing significant advantages in text-rendering and knowledge-intensive generation scenarios. The model excels at tasks requiring precise semantic understanding and complex information expression, while maintaining strong high-fidelity and fine-grained detail generation. Beyond text-to-image, GLM-Image also supports image editing, style transfer, identity-preserving generation, and multi-subject consistency.

Problem it solves

  • Text-rendering accuracy – Best-in-class bilingual (English & Chinese) text rendering within generated images
  • Knowledge-intensive generation – Handles complex, information-dense prompts with high semantic fidelity
  • Instruction following – Precise interpretation of detailed, multi-condition prompts
  • Image-to-image versatility – Supports editing, style transfer, identity preservation, and multi-subject consistency in a single model
  • Creative workflows – Suitable for design, publishing, marketing, and content creation at high resolution

Input/Output

  • Input:
    • Text-to-image: Text prompt (English or Chinese)
      • Format: Natural language description
      • Length: Short to highly detailed (long-form prompts recommended for best results)
      • Language: English, Chinese, or bilingual
      • Examples:
        • "A photorealistic portrait of a woman with blue eyes"
        • "一个穿着红色连衣裙的女人站在樱花树下" (A woman in a red dress standing under cherry blossoms)
        • "Modern minimalist kitchen with marble countertops and stainless steel appliances"
    • Image-to-image: One or more reference images + text instruction
      • Format: Image file (e.g., .jpg) + natural language edit instruction
      • Supports: editing, style transfer, identity-preserving generation, multi-image inputs
  • Options:
    • Diffusion steps (5 - 50): Diffusion steps control how many refinement iterations the model uses to generate an image (default: 50)
    • Random seed
    • Output image height & width (default: 1024×1024)
  • Output: Generated image
    • Quality: Photorealistic, high detail, high-fidelity text rendering
      • image

Accuracy & Speed

  • Quality Metrics:
    • Text Rendering (CVTG-2K benchmark):
      • Word Accuracy: 0.9116 (highest among all open-source models listed)
      • NED Score: 0.9557 (highest among all open-source models listed)
      • LongText-Bench (ZH): 0.979 (top open-source score)

      Text-to-Image (DPG-Bench): 84.78 (competitive with leading open-source models)

      Text-to-Image (TIIF-Bench, short prompts): 81.01 (strong performance among open-source models)

Technical Details

Architecture
Hybrid Autoregressive + Diffusion Decoder
Autoregressive Generator
9B-parameter model initialized from GLM-4-9B-0414; expanded vocabulary to include visual tokens; generates ~256 compact tokens, then expands to 1K–4K tokens for 1K–2K resolution outputs
Diffusion Decoder
7B-parameter single-stream DiT architecture; includes a Glyph Encoder text module for accurate text rendering in images
Key Innovation - Decoupled RL (GRPO)
Post-training with decoupled reinforcement learning: AR module receives low-frequency aesthetic/semantic feedback; decoder receives high-frequency detail/text-accuracy feedback
Training Data
Large-scale image-text pairs; multilingual (English & Chinese); diverse content including photorealistic, artistic, and text-rendering examples
Inference Settings
Guidance scale: 1.5 (default); Height/Width: 1024×1024 (must be divisible by 32); Precision: bfloat16; Requires 80GB+ single GPU or multi-GPU setup
Supported Frameworks
Hugging Face transformers + diffusers; SGLang

Compliance & Provenance

Provider
Open-source (Zhipu)
Provider type
Specialized
License
MIT
EU AI Act risk class
Limited Risk
Art. 50 transparency
Required — outputs are marked. See AI Policy §2.
Region availability
Available globally
Training data summary
Provider model card / training data

For more on how we classify models and mark outputs, see our AI Policy.

Model Source