Why Did an AI Edit Change Something You Wanted to Keep?

You asked for a blue background. The model also changed the person's shirt and softened the face. Your next prompt says “do not change anything else,” but the result drifts again. What should you change next?

Stop adding instructions for a moment. Separate three possible causes: the source does not show enough detail, the request contains competing goals, or the task requires exact preservation that generative editing cannot provide.

Start with a clean comparison

Put the original and the failed result beside each other. Name the error precisely: “left lens became round” is more useful than “looks wrong.” Identify whether the problem is a changed subject, altered composition, or unwanted treatment.

Do not make the failed image your new master reference. If it already contains the wrong glasses, the next edit may preserve those glasses faithfully. Return to the approved source for each controlled attempt.

Test one small change in DrawNew

Open AI Image Editor, select an editing-capable model such as GPT Image 2, and upload the original. Keep the ratio close to the original frame. Remove extra style instructions and try only the background change:

Edit the uploaded photo. Change the background to plain muted blue. Keep the person's face, hair, glasses, expression, clothing, pose, and framing. Keep the existing lighting and natural skin texture. Do not introduce new objects or lettering.

Check the credit estimate first. This deliberately modest brief helps you see whether the problem persists without a new outfit, angle, and lighting setup competing for attention.

Follow the failure, not a longer prompt

What changed	Inspect before the next request	Narrow the next attempt
Face or glasses	Source sharpness and occlusion	Keep pose and lighting unchanged
Product outline	Whether the entire object is visible	Remove camera-angle changes
Label or small print	Readability in the source	Avoid “redesign,” “enhance,” or restyling the product
Framing	Source ratio versus requested ratio	Use a closer ratio and specify the crop
Material color	Colored light and style language	Remove strong palette changes around the item

If the source itself hides the critical feature, a more detailed instruction cannot provide the missing visual evidence. Use the reference preparation guide to decide whether to replace the source.

Are your instructions in conflict?

“Keep the outfit unchanged” conflicts with “make this look like a formal business portrait” if the source shows casual clothes. “Keep every detail” conflicts with “simplify into flat illustration.” State which requirement wins.

For a portrait, keep the clothing and change only the background first. For art conversion, list the defining features to preserve and permit less important details to simplify. The photo-to-art article shows that distinction.

Would a different model help?

You can compare Nano Banana Pro with the same source and brief. Keep a comparable supported ratio and output size. Record what changed, the credits spent, and whether the result is usable.

A single successful sample does not establish that one model will preserve everything in later requests. Use the comparison workflow if you need a repeatable decision for a series.

Know when to switch workflows

If an exact face, logo, or product outline must remain pixel-identical, a conventional selection, mask, or composite can offer control that a generative prompt does not. Do not assume DrawNew provides a pixel-lock or identity-lock control for this workflow.

Stop when the remaining requirement is better handled directly: place the original product over a new background, set exact text in a text layer, or use a simple crop. More generations are useful only when they address a specific unresolved visual choice.