ChatGPT makes and edits pictures right inside a chat. Since Images 2.5 arrived on September 8, 2026, OpenAI says edits change only what you ask for and keep the rest of the image intact, and earlier edits tend to stick across a long conversation. Images 2.5 rolled out to all ChatGPT tiers on desktop, mobile and web. (OpenAI: Introducing ChatGPT Images 2.5 (opens in a new tab))

Images 2.6 followed on November 4, 2026. It adds animated GIFs and reference images you save once and reuse, and it is available across all subscription tiers on desktop, mobile and web, according to OpenAI's announcement and RuntimeWire. (OpenAI Developer Community: Introducing ChatGPT Images 2.6 (opens in a new tab); RuntimeWire (opens in a new tab))

The short version of this guide: describe the picture, then fix it in small steps. Say what must stay the same, change one thing per message, and use a sketch when words are not enough. The prompts below are ready to copy. (OpenAI: Introducing ChatGPT Images 2.5 (opens in a new tab); DataCamp (opens in a new tab))

Start with a clear first prompt

Tip: Write the subject first, then the setting, then the style and light. You do not need a long paragraph. A clear sentence or two beats a list of twenty adjectives, because every extra word is one more thing the model has to balance.

Prompt 1: A clean first image

A close-up photo of a ceramic coffee mug on a wooden kitchen table, morning sunlight from a window on the left, soft shadows, calm and warm.

OpenAI's own example for Images 2.5 shows the style it handles well: concrete details about what a person wears and what surrounds them. Its sample prompt for a 1980s portrait begins with a smiling man with curly hair in a teal, navy and pink windbreaker and a gold chain, then adds the lights and a boombox. (OpenAI: Introducing ChatGPT Images 2.5 (opens in a new tab))

Edit in small steps

Editing is where the tool is strongest. DataCamp's hands-on test found that local edits, such as a new background, a changed line of text or one swapped object, work best, and that the rest of the picture holds up when the change is small and targeted. Moving several objects at once or re-posing a person can break consistency and leave artifacts outside the edited area. (DataCamp (opens in a new tab))

1. Upload your photo or start from a generated image

Add the picture to the chat, or make one with a first prompt.

2. Name what must not change

If a face, a logo or a product shape matters, say so in the same message as the edit.

Prompt 2: Keep the person, change the scene

Change only the background to a quiet beach at sunset. Keep my face, hairstyle, clothes and pose exactly the same.

3. Ask for one change per message

Treat each follow-up as its own small request. If the date on a poster is wrong, ask to change that date, not to redo the whole poster.

Prompt 3: One fix at a time

Change the date on the poster to [NEW DATE]. Do not change anything else.

You can also point at the exact spot. OpenAI says Images 2.5 lets you place comments directly on an image for focused edits, so the model knows which area you mean without a long description. (OpenAI: Introducing ChatGPT Images 2.5 (opens in a new tab))

Use @Sketch and templates

Type @Sketch in ChatGPT to draw directly in the chat and use the drawing as a guide. OpenAI says adding a style description helps. DataCamp found a sketch most useful for spatial details, such as where two characters stand, that are awkward to put into words. You do not need to draw well. (OpenAI: Introducing ChatGPT Images 2.5 (opens in a new tab); DataCamp (opens in a new tab))

Prompt 4: Sketch to finished picture

@Sketch Turn my drawing into a children's book illustration: a small fox on the left, a big tree on the right, soft pastel colors.

Templates give you a starting format such as Poster or Merch. You then add the content, the design elements and the style, OpenAI says. They are good for repeat formats, though DataCamp notes they leave less room for open-ended ideas. (OpenAI: Introducing ChatGPT Images 2.5 (opens in a new tab); DataCamp (opens in a new tab))

Prompt 5: A poster with placeholders

Poster for a neighborhood yard sale. Friendly, bright colors, hand-drawn feel.
Title: Yard Sale
Date: [DATE]
Address: [ADDRESS]
Leave space at the bottom for a small map.

Tip: Mark anything you do not know yet with brackets, like [DATE]. DataCamp's tests suggest giving the event details and the mood, then marking unknowns as placeholders, so you can fill them in later with a one-line edit.

Reuse faces and make GIFs

With Images 2.6 you can save reference images of yourself, your pets and recurring characters and reuse them without uploading them again, OpenAI says. You can also turn a prompt or an image into an animated GIF, then adjust the speed and edit individual frames before you share it. (OpenAI Developer Community: Introducing ChatGPT Images 2.6 (opens in a new tab); RuntimeWire (opens in a new tab))

Prompt 6: A reaction GIF

Make a short looping GIF of my saved cat character waking up, stretching and yawning. Simple, clean background.

What it still gets wrong

Text is still a weak spot, especially small or dense type. Hair edges in portrait edits can look rough, and results can still have a recognizable AI look, according to DataCamp's review. Every image carries C2PA metadata and an invisible watermark that marks it as made with OpenAI tools, the company says. (OpenAI: Introducing ChatGPT Images 2.5 (opens in a new tab); DataCamp (opens in a new tab))

Tip: If a long edit starts drifting, do not keep stacking fixes. Go back to the last good version, restate what must stay the same, and make the next change on its own.