The AR + Diffusion Hybrid Explained (With Diagrams)

GLM-Image uses autoregressive planning for layout + diffusion decoding for pixel fidelity. Here's the intuition, diagrams, and what it means for text rendering.

The intuition: "plan first, render second"

GLM-Image's core design is:

• Autoregressive (AR) stage: generates a compact plan of the image in tokens
• Diffusion decoder: converts that plan into high-fidelity pixels (Z.ai)

This is one reason it can keep layout + typography more consistent than diffusion-only approaches.

Diagram 1: What diffusion-only does

Text prompt
   |
Noise  -> denoise -> denoise -> denoise -> image
         (20–50 steps)

Diffusion denoises the whole canvas repeatedly. Great for texture, weaker for exact letter shapes.

Diagram 2: What GLM-Image does

Text prompt
   |
[AR Planner] -> "layout + meaning tokens"
   |
[Diffusion Decoder] -> pixels

The AR stage is based on a large model initialized from GLM-4-9B-0414 (9B params), and the decoder is a 7B DiT-style diffusion module. (Z.ai)

The token story (why "256 → 4K" matters)

GLM-Image first generates ~256 tokens, then expands to 1K–4K tokens, which correspond to higher-resolution outputs. (GitHub) That expansion is a big part of why it handles complex structured content (posters, menus, infographics).

Why this helps with text

Text rendering is a global constraint:

• letter consistency across a word
• alignment across a column
• spacing across the layout

Planning in tokens first makes those constraints easier to satisfy than trying to "emerge" them from denoising noise.

Practical takeaway for prompt writers

Describe layout zones explicitly:

• "top headline"
• "center hero image"
• "bottom footer bar" …and include the exact required text in quotes. (GitHub)