GLM-Image

lastmod 2026-02-13created 2026-02-131 edits

About

• Autoregressive generator — based on GLM-4-9B (9B parameters). Handles prompt understanding, object relationship reasoning, layout planning, and text processing
• Diffusion decoder — a DiT based on CogView4 (7B parameters). Handles rendering of textures, lighting, and fine details
• The key is a decoupled structure in which the autoregressive model acts as the planner and the diffusion model acts as the renderer.

See also

• Diffusion model