GPT Image 2.5 can build one picture out of several inputs, which is how a group shot, a product placed in a supplied room and a restyle after a reference image all run through the same mechanism.
That mechanism is reference handling rather than a merge button, and the rules are split across three official pages instead of one.
Where the multi-image rule is actually written
The launch notes and the help centre both stay quiet about how many images you may send, and that silence is the first thing to understand.
- The help centre describes uploading “an existing image” and, in the template flow, adding “a reference image” — both singular
- The API reference is the page that states the rule: the edits endpoint “creates an edited or extended image given one or more source images and a prompt”
- Sent as JSON, the images field is an array of objects, each carrying a file_id or an image_url
- For GPT image models that array accepts up to 16 images
- Sent as multipart form data, each image is png, webp or jpg under 50MB
- When you send several images, a mask applies to the first one only
Sixteen is a ceiling rather than a target. Every extra input is another element the model can read the wrong way, and fidelity falls as the count climbs.
Give every input exactly one job
OpenAI describes GPT Image 2.5 as better at working from reference photos “to transform familiar subjects across new settings, visual styles, and compositions”, and the way to use that is to stop treating inputs as interchangeable.
- Image 1 carries identity: the face, the product, the exact object
- Image 2 carries style: line work, palette, lighting, finish
- Image 3 onward carries layout, material or environment
- Name the order in the prompt, as in “Image 1 is the product photo. Image 2 is the style reference.”
- Say what to borrow and what to keep, or the model averages the inputs
- Add position words only when they matter: foreground, background, left, right
The official Codex image generation page gives the same instruction in one line: identify each image by order, then explain how the images relate.
Common combinations and the prompt shape for each
Most multi-image requests fall into a handful of shapes, and each one wants a different sentence.
- Group shot from separate portraits: identity first, one person per input, state the shared setting and lighting once
- Product placed in a supplied room: the product is identity, the room is layout, and the light source belongs in the prompt
- Restyling an existing subject after another image: the subject is identity, the second image supplies style only, and you exclude its content
- Background replacement: one subject image, one background image, plus an explicit instruction not to move the subject
- Flat lay from several objects: enumerate the objects in order and give a count, since counting is a known weak spot
Notice that only one of these asks the model to invent something. The rest are placements, and placements are what reference images are good at.
Why faces and products drift
Consistency across separate generations is a documented limitation rather than something you can prompt away.
- The launch notes claim better subject fidelity, while the API guide still warns that the model “may occasionally struggle to maintain visual consistency for recurring characters or brand elements across multiple generations”
- input_fidelity controls how much effort the model spends matching facial features, and it defaults to low
- Raising it to high costs more and takes longer, so spend it on the identity image rather than on every input
- Keep camera and crop wording identical between rounds, because changing two things at once hides which one moved
- Once a frame is close, edit it instead of regenerating it
- Two people per pass is a sensible default when faces have to stay recognisable
The edit-first habit is the same one that keeps a character stable, which we cover in GPT Image 2.5 character consistency.
What the launch notes do and do not promise
It is worth separating what OpenAI published from what the surrounding coverage assumes.
- The launch notes promise fidelity from reference photos and say reference-led workflows became more reliable for API teams
- They never state a maximum number of inputs
- They never describe how to label images inside a prompt
- They never claim a face will be reproduced exactly
- The four API limitations still apply to combined requests
Anything beyond that list is third-party testing rather than an official claim, including the reported numbers about how many faces line up.
What to do when a combined frame misses
Two of the four documented limitations hit multi-image work hardest.
- Composition control: “the model may have difficulty placing elements precisely in structured or layout-sensitive compositions”
- Latency: “complex prompts may take up to 2 minutes to process”
- Reduce the request to one change per round instead of rewriting the whole brief
- Drop an input rather than adding another when the result turns muddy
- Reserve zones in words before adding people, so each subject has an address
- Reuse the reference order that worked, since the order is doing part of the job
For a small correction inside an otherwise good frame, comment on the spot instead of resending everything, which is what GPT Image 2.5 comment editing is for.
A worked example: a person and a product in one frame
Combining gets hard as soon as two inputs both matter. Take a product shot: a hand holding a coffee tin against a plain wall, where the tin comes from a supplied product image and the hand is described in text.
- Give each input one job and write the jobs down before you send anything: the product image supplies the tin, the text supplies the hand and the wall
- Name the junction explicitly: “The hand grips the tin at the middle of the body, fingers wrapping the front edge.”
- State the lighting once and apply it to both inputs: “One soft source from the upper left, the same falloff on the hand and on the tin.”
- Send a first attempt, then change one thing only: the grip, the angle or the light, never all three
- Check the product’s own details against the supplied image, because logos and label text are the first things to drift
The junction sentence is the part people leave out, and it is the reason hands end up floating beside an object instead of holding it. Both inputs are plausible on their own; that sentence is what binds them.
The related problem of holding one subject stable across attempts is covered in the GPT Image 2.5 character consistency guide.
Frequently asked questions
How many images can GPT Image 2.5 combine at once?
The API reference allows up to 16 input images on the edits endpoint. ChatGPT publishes no equivalent figure, so treat the interface as a few references at most and confirm the behaviour in your own client.
Does combining images work in ChatGPT, or only through the API?
Both work, but only the API documents a number. In ChatGPT you attach a reference image and describe the change, and the official pages describe that flow in the singular.
Can I combine two photos and edit one zone at the same time?
Yes. Send the inputs, then narrow the instruction to a zone. Selections are not pixel-exact, so expect the edit to reach slightly beyond the area you marked.
Why did the second person come out looking like the first?
The model averaged the references. Give each input a distinct job, name them by order in the prompt, and keep the identity image first.
Combining images is mostly bookkeeping: fewer inputs, one job each, and one change per round.