How to Edit Part of an Image with AI (Mark It, Then Describe It)

When you edit an image with AI, or try any AI photo changer, the hard part is rarely the change itself. It is telling the model where to change it. Ask for "a green pillow" in a room with four pillows, and you may get the wrong one, or all of them.

The fix is to point. On CVY.AI you can draw on your image, circle the thing you want changed, and then describe the change in words. The model gets both: where to look, and what to do there.

This guide walks through one real edit from start to finish. We use an image of a living room, change two things in it, then go back and remove an object from the result.

What you need

• The CVY.AI image generator in Image to Image mode, with your image added as reference image 1. Marks only work on the first image.
• GPT Image 2.6 or GPT Image 2.5, Flare or Sunburst.
• Quality turned on. Any level works; we used Standard.
• Animated GIF turned off.

Once an image is in slot 1, a small pen ① appears on it. Pick the model ② and turn on Quality ③.

If you draw marks first and the model can't use them, the generator tells you and offers a Switch button, so you don't have to go looking.

Step 1: Open the marking tools

Click the pen on image 1. The preview on the right turns into a canvas with your image and a toolbar.

• ① Pen draws freehand, Box and Circle draw shapes, Arrow points at something. View lets you move around without drawing.
• ② Three line widths: thin, medium and thick.
• ③ Undo takes back the last mark, Clear removes all of them.

Use + and − at the top right to zoom in on small objects. All marks are red, so pick images where the area you mark isn't already bright red.

Step 2: Mark the area and describe the change

We wanted two changes: swap the yellow pillow for a green one, and replace the little side table with a plant. So we circled the pillow and drew a box around the table.

Then we wrote the prompt. The trick is to refer to each mark by its shape:

Change the circled yellow pillow to a dark green velvet pillow, and replace the boxed side table with a small potted plant.

The prompt ① is where the words go. When you have marks, its placeholder reminds you to say what should go inside them. Press Generate.

Step 3: Check the result

Both marked areas changed. The pillow is green velvet, sitting where the yellow one was. A plant in a white pot stands where the table stood. The rest of the room, including the lamp, the other pillows, the picture and the floor, looks the same.

Under the result there are two ways to keep going:

• ① Mark what to change opens the result on the marking canvas, so you can edit it again.
• ② Continue with… has other next steps, such as using the result as a plain reference image, removing the background or animating it.

Step 4: Remove an object from the photo

Now that the room has a plant, the floor lamp on the left feels like one thing too many. Removing an object from a photo works the same way as changing one: mark it, then say remove.

We clicked Mark what to change under the result and used the Pen ① to draw a loop around the whole lamp, from the shade to the feet. Then we clicked Use these marks ②.

The result becomes reference image 1, with the red loop shown on its thumbnail ①. Your earlier references are swapped out, so this edit starts from the new picture. The prompt box is empty again ②, ready for the next instruction.

The prompt could hardly be shorter:

Remove the circled floor lamp.

The lamp is gone, and the wall and baseboard behind it are filled in cleanly. Look closely at the floor in the corner, though. The lamp's plug and the cable along the baseboard went too, even though they sat just outside our loop. The model treated them as part of the lamp. Here that is what you would want, but it shows that marks are a strong hint, not a hard edge.

From the original image to here took two edits and two short prompts.

How to write the prompt for a marked edit

• Name the shape. "The circled pillow", "the boxed table", "the thing the arrow points to". If you use two kinds of mark, the model can tell which instruction goes where.
• Say what it becomes. "A dark green velvet pillow" works better than "change the color". For removal, "remove" is usually enough. The model fills the gap from what is around it.
• One sentence per change. Two marks, two clear clauses.
• Mark a little wider than the object. Our lamp loop left some space around the shade and the feet.
• Check the edges. Compare the result with the original at full size before you download it.

What about inpainting and masks?

If you have used Stable Diffusion or Photoshop, you know inpainting. You paint a mask over an area, and only the masked pixels are redrawn. GPT Image models also accept a mask, but it doesn't behave like classic inpainting.

OpenAI's image generation guide says masking with GPT Image is entirely prompt-based, and the model may not follow the mask's exact shape. But when we tested it, we found plenty of problems with this approach, mainly that the masked area comes back black.

On the OpenAI developer forum, other developers also report a solid black patch where the mask was. Some found that a fully opaque black-and-white mask, black where the image should change and white where it should stay, may avoid the black patch. But it still doesn't redraw only the masked area. The model treats the mask as guidance, so changes can spill past the edge and areas outside it can shift a little. That approach also contradicts OpenAI's own documentation, which says the transparent parts of the mask are the ones to edit. Replies from OpenAI staff on the forum don't match the documentation either, and third-party platforms each give different advice on how to use masks.

So a mask doesn't give you the hard boundary that inpainting tools do. Marks give the model the same kind of hint, but you draw them right on your image and can see exactly what you are pointing at. You can also combine a pen loop, a box and an arrow in one edit, and refer to each in your prompt.

A few more tips

• Only image 1 can be marked. If your marked image ends up in another slot, you'll see a Move to first button.
• Several changes at once is fine. We changed two things in one go. If one change keeps failing, do it on its own.
• Edit the result, not the original. Each edit builds on the last one, as in step 4, so you don't lose changes you liked.
• Pick 2K or 4K if you need a larger file. The result is a new image at the resolution you choose.

FAQ

Can I edit just one part of an image with AI?
Yes. Add the image as reference image 1, circle or box the part you want changed, and describe the change. The rest of the image is kept as close to the original as the model can manage.

Which models support marks?
GPT Image 2.6 and GPT Image 2.5, both Flare and Sunburst, with Quality on.

Can I remove an object from a photo this way?
Yes. Draw a loop around the object and write "Remove the circled …". The model fills the space with what should be behind it.

Will everything outside my marks stay exactly the same?
Usually close, but not pixel for pixel. Marks guide the model; they don't lock the rest of the image. Check the edges and small details before you use the result.

Ready to try it? Open the CVY.AI image generator, upload an image, and click the pen on it.