Generate an image using an AI image generation model: text-to-image, or image-to-image with reference images. The generated image is attached to your conversation as a native image, so you see it without emitting an [IMAGE:path] marker; the output still carries the marker so the image reaches the user.
The account's chosen model is tried first; when it cannot produce the picture or take the request at all — gone from the catalog, cannot make pictures, cannot take that many references, or cannot express the requested size — the account's other models are tried in turn, and the request succeeds once one serves it. When a fallback served it, the result names the model that produced the picture and what happened with the account's own model (what it answered, or that it was never sent and why); you MUST tell the user both when you send the picture.
Each reference must be PNG, JPEG or WebP. References over 1.5 MB are compressed (downscaled and re-encoded as JPEG), and several may be compressed further to keep the request within the ~2 MB body budget. Local paths are accepted ONLY from the workspace `uploads/` (received attachments) or `generated/` (previously generated media); any other is rejected. Missing, unreadable or undecodable references, more than the provider's 16, or reference material over the total input cap are hard errors — fix or remove them and retry.