Generate an image using an AI image generation model. Supports text-to-image and image-to-image (when reference images are provided). The generated image is automatically attached to your conversation context as a native image, so you can see what you generated without emitting an [IMAGE:path] marker. The tool's output still carries the marker so the image is delivered to the user.
Reference image limits: each reference must be a PNG, JPEG, or WebP file. References larger than 1.5 MB are automatically compressed (downscaled and re-encoded as JPEG) to fit, and multiple references may be compressed further so the total request stays within the ~2 MB request body budget. Local paths are accepted ONLY from the workspace `uploads/` directory (received media attachments) or the workspace `generated/` directory (previously generated media); any other local path is rejected. Missing, unreadable, or undecodable references are hard errors — fix or remove them and retry.