Without the image, “the cup looks wrong” forces ChatGPT to guess which cup, view, and wrongness you mean. With the failed image attached, you can point to an observable relationship: the handle joins too high, the rim is elliptical in the wrong direction, or the logo field wraps around the curve. The attachment establishes the current state; your words define the diagnostic target.

Do not hand the model a verdict without evidence. Replace “cheap,” “off,” or “not realistic” with a feature that could pass or fail: the ceramic glaze has plastic-looking uniform highlights; the contact shadow floats two centimeters below the base; the left margin is narrower than the right; the rear strap duplicates. One visible diagnosis is more actionable than a paragraph of disappointment. Mark the failing region on the frame, then name the intended geometry, lighting, texture, or spacing in the next image prompt.

OpenAI's ChatGPT image guidance recommends describing both what should change and what should stay the same. Put them in separate sentences. First: “Change only the front label area so the rectangle lies flat on the visible face with straight vertical edges.” Second: “Keep the bottle silhouette, camera, cap, liquid level, light, shadow, background, crop, and empty space unchanged.” The separation makes the boundary easy to inspect.

If you attach more than one reference, name each role. The failed image is the composition and identity reference; a second image may define label layout or surface finish. Do not ask both images to define everything. OpenAI's guidance explicitly recommends identifying multiple references and explaining their relationship. Role labels prevent style from quietly replacing geometry or content.

A bounded request can still cause adjacent changes. After ChatGPT returns the edit, cover the repaired region and compare the rest of the frame with the original. Did the cap move? Did the bottle grow? Did the shadow rotate? Did the right-side copy field gain a prop? Only after the protected area passes should you uncover and judge the repair itself.

If the edit fixed the label but altered the silhouette, do not build on it. Return to the attached failure, tighten the locked list, and reduce the change. If the same boundary fails repeatedly, isolate the region more explicitly or make a fresh generation with the approved image as a composition reference. The failed image remains useful because it anchors the known state.

Before prompting: Is the failed image attached? Can you name one failing object, region, or relationship? Is the replacement described with shape, position, light, or spacing rather than taste? Have you assigned every additional reference one distinct role? Does the locked list include camera, crop, object count, and the nearest passing details?

After editing: Did the named region change in the requested way? Did any locked relationship move? Are faces, hands, product geometry, text fields, reflections, and contact shadows still consistent where relevant? Can you state the result as pass or fail without adding a new aesthetic preference? If not, the diagnostic target was still too broad.