ChatGPT Prompts for Photos

What ChatGPT needs before it touches your photo

Upload a selfie, ask for «a professional headshot», and you'll get one. Whether it's still you is another matter. ChatGPT treats every detail you didn't mention as free to redraw, and a face is a lot of details. A photo request gets dependable once it covers five things.

1. Which file. Open with «Use the photo I attached», so the job is an edit of your picture from the first word.
2. The lock. Name what may not change: face, skin tone, hair, glasses, body shape. For a product, the shape, colors and label. Add «no slimming, no skin smoothing» to anything with a face, since a portrait request invites retouching.
3. The change, in visible terms. «Make it look professional» isn't something a camera can do. «Light gray backdrop, soft light from the left» is.
4. The frame. Head and shoulders, waist up or full length, and a ratio for where it's going: 1:1 for most profile photos, 4:5 for an Instagram portrait.
5. A cleanup line. No text, no watermark, no extra people, or a logo may turn up on your shirt.

Put together, a request you can type from memory:

Use the photo I attached. Keep my [face, skin tone, hair, glasses]
exactly as they are. No slimming, no skin smoothing.
Change only [the one thing, in plain visible terms].
[Framing], aspect ratio [ratio].
No text, no watermark, no extra people.

Each prompt on this page is that request worked out for one job. Swap the bracketed words and leave the lock alone.

What to upload

The prompt can only keep what the photo shows. A small, dark or blurry face gets rebuilt from guesses, which is how a result stops looking like you.

• Your face large in the frame. A selfie at arm's length gives the model more of you than a group shot cropped down.
• Even light, eyes visible. Hard shadows and sunglasses hide exactly what the lock asks ChatGPT to keep.
• Whatever the prompt mentions. An outfit change needs the outfit in the picture; a hair color preview needs the hairline and the ends.
• The original file. Screenshots and messenger copies are compressed. OpenAI's help center lists a limit of 20MB per image as of October 2026, so a full-size phone photo is fine.

One rule of ours: edit photos of yourself, or of adults who've said yes. None of these prompts are written for pictures of children.

Which prompt for which photo

Check the last column before you upload.

Keeping a face or a product the same across turns

By the fourth message you've changed the shirt, the backdrop and the light, and the person in the picture looks like your cousin. Each reply is drawn from the last one, so small slips add up; OpenAI's prompting guide for its image models warns that repeated edits can change details you meant to keep. Four habits hold the line.

1. Repeat the lock in every message. The guide's advice is to restate critical constraints when a result drifts, so paste your lock line into each follow-up.
2. Ask for one change per message. If two things change and the face slips, you can't tell which request caused it.
3. Edit from the best version. Don't correct a drifted result on top of itself. Go back to the last image that looked right, or to your upload, and make the next change there.
4. Give each upload a job. With two files, say which is which: «Image 1 is me, image 2 is the jacket. Put the jacket from image 2 on me and keep my face from image 1 unchanged.»

A product follows the same rules. When a logo has to stay exact to the pixel, the guide's answer is plain: paste the approved edit back into the original in an image editor, because prompting alone won't guarantee it. That's why Product Photo From Snapshot changes everything around the item and nothing on it.

One headshot, three turns

A short session for a profile picture. We wrote these messages to show the method; no model output is quoted.

Turn 1

Use the photo I attached. Keep my face, skin tone, hair and glasses
exactly as they are, with no slimming and no skin smoothing. Turn it
into a profile headshot: light gray backdrop, soft light from the
left, a plain navy shirt. Head and shoulders, aspect ratio 1:1.
No text, no watermark.

Say the result is close, but the shirt looks like satin. Don't rewrite the prompt.

Turn 2

Keep my face, skin tone, hair and glasses exactly as in the last
image. Change only the shirt fabric to matte cotton.
Everything else stays as it is.

Now you want a warmer version. Light falls on the face, so this is where the lock matters most.

Turn 3

Use the image you just made. Keep my face, skin tone, hair, glasses
and shirt exactly as they are. Change only the backdrop to warm beige
and the light to late-afternoon window light from the left. My skin
tone stays as it is under the new light. Aspect ratio 1:1.

The lock appears in every turn, nearly word for word, and each turn changes one thing. Turn 3 also names its starting image; if turn 2 had gone wrong, you'd point it at the first headshot. If the face still wanders, open a new chat with your original photo and put everything you've learned into one message. It's often quicker than a fifth correction.

Follow-up lines that fix an edit

When a result is close, one line in the same chat beats starting over.

Where ChatGPT photo edits go wrong

A tight prompt prevents most bad results. These are worth checking for anyway.

• Group photos. Every face is another chance to drift. Group Photo Cleanup locks every person and works only on the background, but check each face at full size before you share it.
• Cartoon versions change you on purpose. A 3D or anime look needs bigger eyes and simpler shapes, so both style prompts aim for «recognizable». A stricter lock fights the style.
• Sharpening guesses. Blurry Photo Sharpen tells the model not to invent detail, yet on a badly blurred face any recovered eyelash is partly made up. Fine for a memory, wrong for anything that has to prove what someone looks like.

What changes the result in ChatGPT

From OpenAI's help center, as of October 2026. Treat them as a snapshot.

• The model. Creating and editing runs on ChatGPT Images 2.5, listed for every tier, free included, on the web, iOS and Android. Images with thinking is listed for Plus, Pro and Business. The DALL·E GPT has been retired, so a guide that sends you there is out of date.
• Two ways to edit. Describe the change in the chat, or open the image, choose Select, paint over one area and describe the change for that part. OpenAI says highlights aren't always precise and an edit may reach past the area you marked, so compare the edges.
• Aspect ratio. Any ratio works, from the aspect-ratio picker or written in the prompt as these blocks do. The editor can also regenerate a finished image at a new ratio.
• Transparent backgrounds. ChatGPT can make a background transparent when asked. For stickers you'll place on a design, put «transparent background» in Selfie Sticker Pack where it names a color.
• Marks on the result. Images made with ChatGPT carry Content Credentials metadata and an invisible SynthID watermark. A screenshot can strip the metadata; the watermark is built to survive light edits.

Prices and plan limits are in our ChatGPT review. To build a photograph with no upload at all, start from the AI photo prompts; for more headshot styles, see the headshot prompts. If you edit in Gemini, use the Gemini photo prompts.