Three useful pictures can make an unhelpful prompt. One shows the object you want, another has the right lighting, and a third suggests a composition. If you upload them with “combine these,” the model has to decide which properties belong in the result. A shape from the lighting reference may arrive along with the light.
GPT Image 2 reference images work better as a brief when you specify what each one contributes. Start with the desired result, assign the uploads their roles and remove any picture that creates a conflict you do not need.
Choose GPT Image 2 reference images that do different jobs
OpenAI's image-edit reference supports multiple input images for GPT Image 2. A tool built around the model may expose fewer upload slots or handle their order differently. Check what the particular interface accepts before preparing a large collection.
For a simple exercise, imagine an original orange desk lamp photographed against a plain wall. That photograph is the subject reference. A second picture of a sunlit desk can describe the lighting and surroundings. A rough drawing can serve as the composition guide, showing where the lamp should sit and where empty space should remain.
Choose the lamp photograph for visible construction details: the shade, stem, base and switch. Choose the desk picture for the light direction and general setting. The rough drawing should be simple enough that it cannot reasonably be confused with a competing finished object.
There is no obligation to use all three. If the desk image already provides the right arrangement, the sketch may add nothing. If it contains a different lamp, consider finding a cleaner setting reference or stating explicitly that its lamp is not part of the brief.
A style reference deserves the same care. Decide whether you want its colour palette, illustration technique or contrast. The phrase “use this style” can ask for more than you intended, including textures or proportions that do not suit the subject.
Keep references that answer a specific question. A folder of attractive images is useful for browsing, but a generation request needs a clearer set of instructions.
Write the relationship between the images
Use the labels the tool presents, or describe each image unmistakably if it has no labels. Do not assume that filenames are visible to the model. Confirm the upload order after all attachments have loaded, particularly if the application reorders them.
For the lamp exercise, the brief could read: “Use the orange desk lamp in image 1 as the subject. Use image 2 only for the warm light from the left and the wooden desk setting. Use image 3 for placement: the lamp occupies the right side, with empty space on the left. Keep the lamp's shade, stem and base proportions from image 1.”
Now a reviewer knows which photograph to use when checking the lamp and which to use when checking the light. The composition sketch has a limited job too: placement, with no claim over the object's appearance.
Make priorities explicit where the inputs disagree. If the room reference shows cool overhead lighting but the written brief calls for warm light from the left, resolve that conflict before generating. A reader should not need to guess which instruction you consider more important.
Specify the intended output as well. Combining references into one new scene is different from making a contact sheet that shows each source separately. Multiple reference images are also different from requesting several output candidates. An interface may support one operation while offering limited control over the other.
Avoid asking the model to fill gaps that your own team has not decided. If the lamp's hidden side matters, provide a suitable photograph or mark that area as unresolved. A generated view may be visually coherent without representing the actual object accurately.
After the first result, compare it with the brief in the same order: subject, setting, composition. Check the lamp's construction before spending time on the warmth of the desk. A lovely room is of little use if the item the picture was meant to depict has changed.
Remove a reference when it creates a conflict
Suppose the generated lamp has acquired a curved base from an object in the desk photograph. First verify that the prompt assigned the references correctly and that the app retained the intended upload order. Then test the subject photograph with a written description of the setting, leaving the desk image out.
This does not prove exactly how the model used either picture. It does give you a practical comparison: whether the unwanted curve appears when that reference is absent. Keep the size, prompt wording and other inputs as stable as the tool allows while checking the change.
If the shape improves but the lighting becomes less convincing, you have a choice. Find a cleaner lighting reference, describe the light more precisely or finish the lighting in an image editor. Uploading several more room pictures may make the source of the problem harder to identify.
When trying the same brief with Nano Banana Pro, carry over the reference roles, not unexplained interface settings. Verify how the second tool labels and accepts its inputs. Compare the results against the intended lamp and layout rather than assuming that the same number of attachments means the same behaviour.
For an image that must preserve an exact product component, ordinary compositing may be preferable. Reference-based generation offers visual guidance; it does not provide the same control as keeping an original object on a separate, untouched layer.
Save the final reference set alongside the prompt, with a note identifying the subject, lighting and layout sources. When someone returns to those GPT Image 2 reference images for a revision, they can replace the room without accidentally replacing the lamp's design.