GLM-Image

Overview

Z.ai built GLM-Image from two models that run in sequence. A 9B-parameter autoregressive generator, initialized from GLM-4-9B-0414, writes a compact encoding of about 256 visual tokens and expands it to between 1K and 4K tokens. A 7B-parameter single-stream DiT decoder then renders the image, helped by a Glyph Encoder module for lettering.

Z.ai aims it at pictures that carry information: commercial posters, popular-science diagrams, multi-panel comics and social media covers. It reports a word accuracy of 0.9116 on CVTG-2K, which asks for several text regions in one image, and 0.952 (English) and 0.979 (Chinese) on LongText-Bench. On general image quality, Z.ai places it level with mainstream latent diffusion models.

The same weights also do image-to-image work: editing, style transfer, identity-preserving generation and multi-subject consistency.

What you can build

• Promotional posters with a headline and a few short lines of copy, checked for typos before use
• Recipe cards and science explainers with labeled steps
• Story comics and product sets that keep one character or item consistent across panels
• Memes and social covers with large captions
• Background swaps and restyled versions of an existing photo

Inputs

Only the prompt is required. Z.ai recommends detailed prompts, and the page's default prompt is a good model: a full paragraph laying out a magazine-style recipe page with a title, seven ingredients and four captioned step photos.

To edit or combine pictures, upload up to 5 images (JPG, PNG, GIF, WebP or HEIC) and describe the change in the prompt.

The other controls are width and height (each up to 2048, default 1024), inference steps (1 to 50, default 30), guidance scale (1 to 20, default 1.5), a seed, and Samples for up to 8 images per run.

Outputs

One image per sample at the size you set. The page's 1024 by 1024 samples include the recipe guide from the default prompt, a street portrait beside a painted wall slogan, and a cluttered-desk meme captioned "I LOVE BEING ORGANIZED" and "in theory."

Recommended settings

• Put every word you want drawn inside quotation marks, as Z.ai's card asks.
• Width and height snap to the nearest multiple of 32 between 512 and 2048 (a requested 1000 comes out at 992), so use multiples of 32 such as 1024, 1056 or 1568 to get the exact size.
• In edits, the output size still comes from width and height, so match them to the source image's proportions.
• Z.ai's examples use 50 steps at guidance 1.5; this page defaults to 30 steps.
• Z.ai suggests expanding short prompts with GLM-4.7 first.

Limitations

• Small text in dense layouts still slips. In the recipe sample the title and headings are clean, but one ingredient reads "Fiuw" for "Flour", a step caption stops at "chlll for 4" and another reads "Rasparties".
• The same sample dropped the "Cooking time" and "Servings" labels from its bottom bar, leaving only "20 minutes" and "8" beside their icons.
• Z.ai reports text rendering for English and Chinese only.

Safety & compliance

GLM-Image is released under the MIT License. Its VQ tokenizer and ViT weights come from X-Omni and stay under Apache-2.0. If you upload a photo of a real person for identity-preserving generation, get that person's consent first.