ChatGPT can edit an image you upload or an image previously generated in ChatGPT. Open the image, describe what you want changed and optionally highlight the relevant area before submitting your instruction. Because selections and generated edits are not perfectly precise, review the entire result before saving it.
The workflow below applies to ChatGPT on the web, iOS and Android. Control placement and wording can vary slightly by platform and interface rollout, but the underlying process is the same. This guide covers modifying an existing image rather than asking ChatGPT to analyze what the image contains.
Open the image you want to edit
You can begin with either an existing image from your device or an image that ChatGPT has already generated. According to OpenAI’s Images in ChatGPT guide, both starting points support image editing.
- For an image on your device: Open a ChatGPT conversation and use the available attachment or upload option to add the image. Once it appears in the conversation, open the image.
- For a ChatGPT-generated image: Open the conversation containing the image and select the image you want to modify.
- Open the available editing interface or continue in the chat with an instruction describing the change.
On iOS and Android, tapping the image opens the mobile image view, where the available editing and area-selection options appear. On the web, open the image from its conversation before choosing the relevant editing option. The exact position of these controls may differ between current interfaces, so use the image-level editing controls shown on your device.
Before proceeding, make sure you opened the correct image and, if several similar versions exist in the conversation, the exact version you want to change. Each edit creates a generated result based on the selected source image.
Describe the change you want ChatGPT to make
You can request an edit in plain language. Identify the element or region to modify, then state the replacement or adjustment you want. If the target is unambiguous, you can submit the instruction without highlighting an area first.
Useful instructions are specific about both location and outcome. For example:
Replace the gray wall in the background with a pale green wall.
Other targeted requests might change an object, remove an unwanted visual detail, adjust text treatment or alter the framing. Instead of asking ChatGPT to “improve the image,” describe the visible change you expect. For example, specify “change the red mug on the left to dark blue” rather than “fix the mug.”
If the image contains repeated objects, identify the target by position, color or relationship to another object. A phrase such as “the small lamp on the right side of the desk” gives ChatGPT more useful guidance than simply saying “the lamp.”
Submit the instruction and allow ChatGPT to produce the edited version. A clear prompt can improve control, but it does not guarantee that every changed pixel will be limited to the named object.
Select a specific area for a targeted edit
If describing the target is not enough, use the image editor’s area-selection option. This is useful when you need to change one object, a small region of the background or another clearly bounded part of the image.
- Open the image in its editing view.
- Choose the available area-selection option.
- Highlight the part of the image you want ChatGPT to modify.
- Confirm the selection if the mobile interface asks you to proceed.
- Enter an instruction explaining what should replace or change within that area.
Treat the highlighted region as guidance rather than a strict editing mask. OpenAI warns that a selection may be imprecise and that the resulting edit can extend beyond the highlighted area. Details immediately around the selection—and sometimes elsewhere in the image—may also change.
For that reason, highlight enough of the relevant object or region to make the intended target clear, but do not assume that everything outside the selection will remain identical. Pairing the selection with a precise written instruction provides more context than relying on the highlighted area alone.
For example, after selecting a flower pot, you could write:
Change the selected flower pot from white ceramic to matte terracotta. Keep the plant, tabletop and surrounding shadows unchanged.
Write a prompt that preserves the rest of the image
A targeted editing prompt should explain what to change and what matters outside that change. OpenAI’s image prompting guide recommends describing targeted revisions and specifying details that should remain unchanged.
A compact prompt pattern is:
Change X to Y while keeping A, B and C unchanged.
Replace those placeholders with visible, concrete details. For example:
Replace the cloudy sky with a clear blue late-afternoon sky. Keep the buildings, camera angle, warm lighting and people unchanged.
Depending on the image, preservation instructions can refer to:
- Composition: Keep the same crop, camera angle and object positions.
- Lighting: Preserve the direction, intensity and color of the existing light.
- People: Keep the same faces, expressions, poses and clothing.
- Colors: Preserve the existing palette outside the edited object.
- Text: Keep specified wording and its placement unchanged.
- Background details: Preserve furniture, buildings, plants or other surrounding elements.
List only the details that are important to the result. An excessively broad instruction can make the main change less clear, while a vague instruction gives ChatGPT more freedom to reinterpret the scene.
Preservation wording improves the direction of the request, but it is not a guarantee. Image edits are generated rather than applied as deterministic layer-based changes, so inspect important faces, text, logos, patterns and small objects after every revision.
Review, refine and save the edited image
When ChatGPT returns the edited version, compare it with the source image. Do not inspect only the selected region, because an edit can affect details outside that area.
Check the following parts of the result:
- The requested element has the intended color, shape, material or appearance.
- The composition and framing still match the original where required.
- Faces, hands, text and small objects have not changed unexpectedly.
- Lighting and shadows remain consistent around the edited area.
- Background details outside the target region are still acceptable.
If too much changed, submit a narrower follow-up instruction. Restate the target and repeat the most important preservation requirements rather than relying on ChatGPT to infer them from the previous request.
For example:
Revise the last edit. Change only the wall color to pale green. Restore the original furniture, window, lighting and framing.
If the first instruction was ambiguous, identify the target more precisely by its position or surrounding details. You can also reopen the preferred image version, select a smaller or clearer region and submit a revised instruction. Results can still vary, so choose the version that best preserves the source rather than assuming the latest edit is automatically the best one.
Once you are satisfied, open the finished result and use the available save control. On a phone or tablet, follow the device prompt that appears after saving. On the web, use the image’s save option and then locate the file using your browser’s normal saved-file workflow. Review the saved image once more to confirm that you preserved the intended version.
Can ChatGPT edit both uploaded and generated images?
Yes. OpenAI documents editing for an existing image you upload as well as an image generated in ChatGPT. Open the chosen image, provide an editing instruction and save the result when you are satisfied.
Can I change only one part of an image in ChatGPT?
You can use the optional area-selection tool to indicate the part you want changed. The selection is approximate, however, and the generated edit may extend outside the highlighted region. Review the whole image afterward.
How do I keep the rest of the image unchanged?
Name the exact element to change and explicitly list important details to preserve, such as composition, lighting, colors, people or text. This improves the direction of the request but cannot guarantee that every other detail will remain identical.