Image generation

An assistant with an image model can produce an image from a description, and edit one you give it. Ask for what you want in the conversation the same way you would ask for text — the image comes back in the conversation, and you can download it or keep working on it.

Asking for an image

Describe the image you want. The more specific the description, the closer the result: subject, style, what should and shouldn’t be in frame. If the first attempt isn’t right, say what to change rather than starting again — the assistant works from the previous image.

Shape, background and file format

An assistant can choose the image’s proportions, its background and the file format it produces, rather than always returning the same square opaque picture. That matters most for the cases where an image has to fit somewhere specific:

• A transparent background gives you a real cutout. Ask for a logo or an icon on a transparent background and you can drop the result straight into a document or a slide without a white box around it.
• The shape can follow the use — a wide banner, a tall poster, a square avatar.
• The file format follows the same logic: a format that supports transparency when the image needs a cutout, a photographic format when it doesn’t.

You don’t have to name any of this. Describe what the image is for and the assistant generally picks sensibly; name it explicitly when the result has to match a specification.

When a request is refused

Not every image request can be fulfilled. Where one is turned down — because of the model’s own content rules, for instance — the assistant tells you why in plain language, rather than reporting an internal error. Read the reason before retrying: a rephrase fixes some refusals and will never fix others.

Editing an existing image

Attach an image to the conversation and ask for the change you want. The assistant works from the image you gave it rather than generating something new from scratch.

Availability

Note

Image generation needs an image model configured for your organization and permitted by the security classification of the Space you’re working in — including the regions that classification allows data to be processed in. The model is chosen under AI setup when you build the assistant. If you don’t see an image option where you expect one, check with your administrator.