How do you edit a photo with AI in Gemini or ChatGPT?
Gemini edits with Google's Nano Banana image models, and OpenAI offers ChatGPT Images on every ChatGPT plan, free included. The steps are the same in both apps:
- Open a chat app that edits. Use gemini.google.com or the Gemini app, or ChatGPT on the web or your phone. Both keep the thread, so a second fix builds on the first.
- Attach the photo. Tap the plus or the paperclip and pick the picture. Gemini only lets adults edit: Google's help page turns editing off for anyone under 18.
- Write the change, then the keep list.
Remove the red car on the left. Keep the people, the light and the framing exactly as they are.OpenAI's own guide says its model treats even a painted mask as guidance, so spell out what to protect. - Fix one thing per turn. If the sky came out right and the face changed, say so:
Keep this sky. Restore her face exactly as it was in the original photo. - Download, then zoom in. Gemini saves at 1K without a Google AI plan and 2K with one. Check hands, edges and any text before you post.
ChatGPT also has a selection tool. Brush over the part you want changed, then type the change. OpenAI's help page warns that "Highlights are not always precise, and edits may extend beyond the area you selected." So check the whole frame, not just the part you painted. Free use is capped in both apps, and the caps each vendor publishes are on our ChatGPT image limit and Gemini image limit pages.
Which AI models can edit a photo you upload?
Only models trained to follow an edit instruction can change your photo. A pure text to image model paints a new picture from words, and the chat window looks the same either way, so it's easy to miss.
OpenAI's model page for DALL·E 3, the model behind ChatGPT's first image tool, lists the image edit endpoint as not supported. OpenAI retired it on 12 May 2026. Black Forest Labs sells editing as its own model, FLUX.1 Kontext, which needs a photo plus a prompt. Its newer FLUX.2 takes up to 8 reference photos through the API. Google's Nano Banana models take up to 14.
OpenAI doesn't publish a per edit photo cap for ChatGPT. Its API guide shows four photos combined in one request, and the API's largest output is 3,840 pixels on the long edge. If you want a picture made from nothing, that's a different job with different winners, covered in best AI image generators.
Whizi follows the same rule. Its photo editor is Nano Banana, one edit takes up to 4 photos, and the result keeps the shape of the photo you sent instead of cropping it square. Pick Flux or Stable Diffusion with a photo attached and Whizi hands the job to Nano Banana. Those two only offer image to image on fal, the host Whizi uses, and that repaints the frame. Where Whizi loses: its editor runs Nano Banana 2 Lite, which Google lists at 1K output only and says isn't designed for several reference photos or repeated edits. A 4K print needs Nano Banana Pro in Google's API. The steps inside Whizi are in generate images in chat.
12 photo editing prompts to paste into Gemini or ChatGPT
Each prompt puts the change first and the keep list second. Swap the brackets for your own details and send it with the photo attached. We didn't benchmark these for this page, so treat each one as a first draft and use the right column to judge the result. Nothing in them is tied to one app, so you can paste them into Gemini, ChatGPT or any chat that edits with Nano Banana.
| Edit | Prompt to paste | Check before you keep it |
|---|---|---|
| Remove an object | Remove the [bin] from the [left] of the photo. Fill the space with the [wall] around it. Change nothing else. | Repeated patterns where the object was |
| Remove a person | Remove the person in the [red jacket] behind us. Rebuild the background they cover. Keep our faces, clothes and poses exactly the same. | The faces of the people who stay |
| Swap the background | Replace the background with [a plain light grey studio wall]. Keep the subject, hair edges and shadow unchanged. Light the subject to match the new background. | Hair edges and the shadow |
| Relight | Relight the photo as warm late afternoon sun from the left. Keep the framing, faces and clothing colours the same. | Skin tone and eye colour |
| Fix exposure and colour | Brighten the shadows and remove the yellow cast so whites look neutral. Do not add, move or remove anything. | Nothing new appeared |
| Restyle | Turn this photo into a [watercolour painting]. Keep the same people, positions and framing. | Faces still look like the people |
| Combine two photos | Put the dog from photo 2 on the sofa in photo 1. Match the light, scale and shadow of the room. Keep the framing of photo 1. | Scale, and where the shadow falls |
| Change a colour | Change the jacket to navy blue. Keep its texture, folds and shadows. Change nothing else. | Colour bleeding onto skin |
| Clean up a product shot | Put this product on a plain white background with a soft shadow under it. Keep the label text, shape and colours exactly as they are. | The label, letter by letter |
| Restore an old photo | Remove the scratches and dust, fix the faded colours and sharpen gently. Do not change the faces or invent details. | Invented details in faces |
| Widen the frame | Extend this photo to a 16:9 frame by adding more [beach] on both sides. Keep the original picture untouched in the middle. | The join on each side |
| Replace text on a sign | Change the sign to read "[new words]". Keep the font, size, colour and angle of the original. | Spelling, which OpenAI lists as a known weak spot |
Two habits are worth building. Name the light when you change a background, because a subject lit from the left on a scene lit from the right looks pasted on. And when an edit goes wrong, don't stack a second fix on top. Go back to the original photo and send a tighter prompt.
What do Gemini and ChatGPT stamp on an edited photo?
Both stamp the same invisible watermark as of October 2026: SynthID, which Google DeepMind built. Google puts it in every image Gemini creates or edits. OpenAI added it next to its C2PA Content Credentials in ChatGPT, Codex and its API, per the system card for its September 2026 image model. Google says SynthID survives resizing, recolouring and compression, though heavy changes can remove it.
| What gets added (October 2026) | Gemini app (Nano Banana) | ChatGPT Images |
|---|---|---|
| Invisible watermark | SynthID in every image created or edited | SynthID, added with the September 2026 model |
| Visible mark | Gemini sparkle, optional since Google announced the Settings, Media watermark switch on 14 August 2026 (not offered where the law requires a visible label) | None mentioned in OpenAI's help pages |
| Metadata | C2PA metadata, per Google (August 2026) | C2PA Content Credentials, lost if you screenshot or convert the file |
| Who can edit | Adults only, 18 or older | Every ChatGPT plan, free included |
| Download size | 1K without a Google AI plan, 2K with one | Not stated in the help page. The API caps the long edge at 3,840 pixels |
Switch off the sparkle and the visible label goes. SynthID stays. Google's Josh Woodward announced the switch on X on 14 August 2026 and said SynthID and C2PA metadata are still used in every file. Anyone can upload a picture to Gemini and ask whether Google AI made or edited it, with about 10 image checks per 24 hours.
Why does Gemini or ChatGPT refuse to edit a photo?
Both refuse edits that could pass for a real photo of a real person doing something they didn't do. OpenAI's system card, dated 8 September 2026, names convincing deepfakes of real people, including political and sexual imagery, as the main new risk. Its safety model reads the uploaded photo and the prompt together to catch malicious edits, then checks the finished image a second time before anything reaches your screen.
The filter isn't perfect in either direction. On OpenAI's own set of prompts written to break policy, GPT-Image-2.5-Sunburst blocked 21.9% and still showed an unsafe image 1.09% of the time. Google turns editing off for anyone under 18, and Gemini can remove an image when its systems flag a possible violation.
Edits to your own everyday photos are the intended use. Google launched its August 2025 upgrade by pitching new outfits, new wallpaper and blended photos where people and pets still look like themselves. If a harmless edit is refused, describe what the pixels should look like, not why you want it.
When does a regular photo editor beat a chat prompt?
A pixel editor wins when every untouched pixel has to stay exactly as it was. Chat editors redraw the picture. OpenAI's API guide says a mask is only guidance and the result may not follow its shape. It also lists text, precise layout and consistency across images as weak spots.
So split the work. Use the chat for big changes: a new background, new light, a stranger removed. Use Photoshop, Lightroom or the editor built into your phone for a crop, an exposure nudge or a small blemish. Do the same for any photo you'll submit somewhere that rejects AI edits.
One-tap AI fixes don't need a chat at all. Google Photos has Magic Eraser, which Google says removes unwanted objects in a few taps. Adobe Express Photos has an AI assistant with one-click buttons for removing distractions and changing backgrounds, plus a box for typed requests. Reach for those when the fix is a common one and you'd rather tap than write a prompt.
- Use a model built for editing: Nano Banana in Gemini or ChatGPT Images, not a text to image model
- Write the change first, then a keep list naming faces, light and framing
- Change one thing per turn and restart from the original when an edit goes wrong
- Zoom in on hands, edges and text before you post the result
- Expect SynthID and C2PA metadata in every Gemini or ChatGPT edit, even with the visible sparkle off
Frequently asked questions
Can ChatGPT edit my own photos?
Yes. Upload the photo on any ChatGPT plan, free included, and describe the change, or brush over one area with the selection tool first. OpenAI warns the selection isn't always precise, so check the whole picture. Edits that could pass as genuine political or sexual images of real people are refused.
Can you edit photos with AI for free?
Yes, in both main apps. ChatGPT Images runs on every ChatGPT plan, and Gemini edits photos without a Google AI plan but downloads them at 1K instead of 2K. Free use is capped in both, and Gemini editing is for users 18 or older.
How do I turn off the Gemini watermark on an edited photo?
Open Gemini Settings, choose Media watermark and switch it off. The change applies to images you make after that, and it removes only the visible sparkle. SynthID and C2PA metadata stay in the file, and the switch isn't offered in countries where the law requires a visible label.
Can I edit photos with AI on my iPhone?
Yes. The ChatGPT and Gemini apps for iPhone take a photo from your library and edit it from a typed request, the same way their websites do. Gemini editing still needs an account holder who is 18 or older, and free use is capped in both apps.
