ChatGPT imagesConsistent charactersStyle references

ChatGPT Image Edits and References: Keep Things Consistent

By Marco Cavazzana · Updated · 5 min read
ShareEmail
Short answer

ChatGPT can edit an image you upload and use uploaded pictures as references for new ones. Attach the image, say precisely what to change and what to keep, and make one edit per message. For a consistent character or style across several images, keep the same reference attached and reuse an identical description in every prompt.

The GPT Image model in ChatGPT does more than generate from text. You can upload a photo or an earlier generation and ask for changes, or attach one or more images and ask for something new in the same style or with the same character. Because the whole thing happens in a conversation, edits build on each other turn by turn. This guide covers how to phrase edits so they stay surgical, and how to use references to keep a character, product or look consistent across a set of images.

Two ways to use an uploaded image

It helps to be explicit about which of two things you are asking for, because the model will otherwise guess.

  • Edit. "Change this image": the output should be the same picture with a modification. Replace the background, recolour the jacket, remove an object, extend the frame.
  • Reference. "Make a new image using this": the output is a fresh composition that borrows something from the upload, such as a character's face, a product's shape, a colour palette or an illustration style.

Say which you mean in the first sentence: "Edit the attached image so that..." or "Using the attached image as a reference for the character, create a new image of...". That single sentence prevents most disappointing results.

Making clean edits

Multi-turn editing is the model's strongest feature, and it rewards discipline.

  1. One change per message. "Make the background a plain pale blue" then, next message, "now add a shadow under the product". Bundled edits are harder to judge and harder to undo.
  2. Name what to keep. "Keep the person, pose, lighting and framing identical." Without this, the model may tidy things you liked.
  3. Describe the target precisely. "The mug on the left of the table" rather than "the mug".
  4. Prefer positive instructions. "Make the wall white" is more reliable than "remove the colour from the wall".
  5. Branch, do not pile. If the fourth edit ruins something, go back to the message with the last good image and continue from there instead of asking for a fix on top of the mistake.

Expect small drift. Even a well-phrased edit may slightly change textures, a face or the exact position of an object, because the model regenerates the image rather than masking a region in the way a photo editor does. For edits where pixel-exact preservation matters, keep the changes small and check the areas you did not ask to change.

Keeping a character consistent

Consistency across a set comes from three things working together: a good reference image, an unchanging written description and consistent framing choices.

  • Build a reference first. Generate your character once, front-facing, neutral expression, plain background, even light. Iterate until you are happy, then save that image. It is your master reference.
  • Attach it every time. Each new image should have the master reference attached and a sentence such as "the woman in the attached image is the same character; keep her face, hair and build identical".
  • Reuse one description. Write a two-line description of the character (age range, hair, distinguishing features, standard outfit) and paste it unchanged into every prompt. Do not paraphrase it.
  • Change one variable at a time. New pose with the same outfit and setting, then a new setting with the same pose. Changing everything at once is where identity slips.
  • Keep faces large enough. Tiny faces in wide shots are the first to drift. Where possible, frame medium or closer.

The same logic applies to products: a master image of the product on a plain background, attached to every scene prompt, with its shape and colour described identically each time. For product work specifically, see ChatGPT images for product mockups. The approach also mirrors how reference workflows work elsewhere; if you use Midjourney, compare Midjourney consistent characters.

Using images as style references

A style reference is an upload you want the model to learn the look from without copying the content. Say so explicitly: "Use the attached image only as a style reference: match its colour palette, line quality and lighting, but do not reproduce its subject." Then describe your new subject fully.

Use your own work or images you have the right to use as style references. Describing a style in words ("flat vector, limited palette of teal and mustard, thick outlines, paper texture") alongside the image gives the model two sources of truth and tends to be more stable than the image alone. Avoid asking for the style of a named living artist.

Settings at a glance

SettingWhat it doesStart with
Attached imageThe picture to edit or referenceOne clean image; more only when each has a clear role
Mode sentenceTells the model whether to edit or create new"Edit the attached image..." or "Using the attached image as a reference..."
Keep clauseLists what must not changePose, face, lighting, framing
Character descriptionFixed text reused in every promptTwo lines, pasted verbatim
OrientationSquare, portrait or landscapeMatch the reference unless you need otherwise
BackgroundScene, plain or transparent PNGPlain for master references

Try it

Using the attached image as a reference for the character, create a new portrait-orientation image. She is the same person: keep her face, short dark curly hair, round glasses and the mustard scarf identical. New scene: she is reading on a train by the window, late afternoon light, soft focus landscape passing outside. Medium shot, eye level, natural photographic style.

Attach your master reference, replace the fixed description with your own two lines, and only change the sentence beginning "New scene". If the face drifts, move to a closer shot or regenerate your master with flatter light and a more neutral expression.

Frequently asked questions

Can ChatGPT edit a photo I upload?
Yes. Attach the image and describe the change, such as replacing the background or recolouring an object, and say what must stay the same. The model regenerates the image with your change, so expect minor variation in untouched areas and keep edits small where precision matters.
How do I keep the same character across several ChatGPT images?
Create one master reference image of the character on a plain background, attach it to every new prompt, and paste the same short written description each time. Change only one thing per image, such as the pose or the setting, and keep the face reasonably large in frame.
Can I use an image as a style reference without copying it?
Yes. Tell ChatGPT to use the attached image only for its style, naming the qualities to match such as palette, line quality and lighting, and describe a different subject. Use images you own or have rights to, and avoid asking for a named living artist's style.
DreamdriveComing to Chrome
Join the waiting list