We use Generate Image to make images. It cannot specify position, size, or typeface. Revising after an image is produced is normal. A short instruction is rejected. A long description is also rejected. An image from the same run has already come out; change a few words and it is rejected again. We keep testing and still do not know how to work with it. About eight out of ten tries are rejected. The error is: Image generation was not performed — the tool call appears to be unintended.
Steps to Reproduce
In Cursor IDE, open Agent chat on a project.
Ask Generate Image with a short instruction. It is rejected.
Ask again with a long description. It is rejected.
One image comes out. Change a few words and ask again. It is rejected.
Error: Image generation was not performed — the tool call appears to be unintended.
Expected Behavior
When I explicitly ask for an image, Generate Image should produce it. Changing a few words to revise that image should produce another image, not reject the request.
Hi @CEO_AdFinTech, thanks for the report and for linking the earlier thread.
This is a known issue we’re already tracking, and I’ve added your report to it. The check looks only at the short image description the agent writes when it calls Generate Image, not at your message. If that description is very short, it gets rejected. That’s also why a small revision often fails: the agent tends to write only the change (for example “make title red”) instead of the whole image description.
Could you try this and tell me whether the rejections stop?
Start a new Agent chat and select Grok 4.7.
Ask for the image and add: “When you call Generate Image, write the full image description in English, at least 20 words, including layout, colors, text and typeface.”
For a revision, ask for the full updated description again with your change included, not only the change.
One more question: do you usually write your prompts in Chinese? Descriptions written without spaces between words are more likely to be rejected right now, so an English description is the most reliable workaround for the moment.
We already sent the full description in English. It was not rejected. The image came out. The same English description and the same layout, sent again, come out different every time. Specifying position, size, typeface, and effect in the description does nothing. The image does not follow them.
Telling us to select Grok 4.7 does not make sense. The chat model does not draw the image. The image model is one you connect on your side. It cannot be chosen in settings. Selecting Grok in the chat does not make Grok draw the image. Grok 4.7 is the most expensive model in the usage pool. Image generation still asks us to keep that model on, and the picture is still drawn by your fixed image model.