How to Combine Multiple Reference Images with AI

Decide whether you need a new composition or an edit to a fixed image

Start with the part of the picture that must remain most faithful.

If you want a new composition, all references can guide a newly generated scene. This suits a character in a new location, a mood-board concept or an illustrated campaign image.

If you must preserve an existing photograph, identify it as the base and request a limited addition or change. For example: “Keep image 1's room, camera angle and furniture unchanged. Add the chair from image 2 beside the window.” Even with that instruction, inspect the whole result for unintended changes. Use a selection or mask when the chosen editor supports it and the change belongs in a specific area.

If exact pixels, product lettering or geometry must survive, a layered image editor may be a better final assembly tool. AI can help harmonize a composite, but a reference is guidance rather than a promise to paste its contents unchanged. The comparison of tools for editing part of a photo explains the control choices in more detail.

1. Give every image one main job

You do not need a large reference collection. Begin with the smallest set that contains the information the finished picture needs.

A style reference and an identity reference can conflict. A highly stylized face may encourage the model to change the original person's proportions; an elaborate background may pull the camera away from the planned crop. State which role has priority.

OpenAI's image-generation guidance likewise recommends assigning roles and describing spatial relationships. That principle travels across tools, but image limits and reference controls depend on the selected product and model.

For a new scene with a known character, use the clearest approved identity image. A front or three-quarter view can make identifying features visible; add another view only when it supplies information the scene needs. For a product, include the angle you want shown rather than expecting the model to reconstruct a hidden label.

2. Prepare a reference set you can explain in one sentence

In the worked example, the brief was simple: place a small courier robot on the right side of a botanical conservatory at blue hour, leaving room on the left for later text.

Three images supplied different information:

1
1. Robot reference: cream body, asymmetric amber and teal eyes, red-bead antenna, mustard backpack and dark-blue hands.
2
2. Conservatory reference: glass roof structure, warm doorway, cool evening light, planting and terracotta walkway.
3
3. Layout guide: a right-side subject and a left-side area reserved for text.

Reference 1 supplies identity. It is not the intended background or final crop.

Reference 2 supplies the setting and lighting. The roof structure and doorway make it more specific than “a greenhouse.”

Reference 3 supplies placement. Its labels should not appear in the finished image.

Your references do not have to share the same visual style, but explain how the final picture should reconcile them. Here, the requested treatment was tactile stop-motion illustration. The subject's identity mattered more than copying every pixel of its neutral background.

3. Upload the references and check their order

Free accounts in the US receive 120 credits per day; daily allowances vary in other regions. You can start exploring image ideas with Seedream 5.0 Lite and Nano Banana. Open Dreamina's AI photo editor and enter AI Image. Add the reference images, choose an available model that accepts your set, and select the output aspect ratio and resolution you need. Model availability and credit requirements can vary by account.

Before generating, check which thumbnail is reference 1, 2 and 3. If the interface reorders them, update your prompt. A perfectly written instruction attached to the wrong image can swap the subject and background roles.

The example used Seedream 5.0 Pro, 16:9, 2K, with three references. These are the settings of this example, not a claim that every model supports the same reference count. The downloaded result was 2560 × 1440 pixels.

Choose the aspect ratio before working on small details. A later change from a square to a wide banner can reposition the subject and use the space you intended for text.

4. Write a prompt that connects the images

A useful prompt names the reference roles, describes the finished scene and protects the important details. “Combine these images” leaves the model to decide too much.

This was the prompt used with the three references:

Use reference 1 only for the courier robot identity, reference 2 for the conservatory architecture and lighting, and reference 3 for layout only. Put the exact robot naturally on the right-side walkway, leave the left third clear. Preserve its asymmetric eyes, red-bead antenna, mustard backpack and blue hands. Match ground perspective and contact shadow. Do not copy guide labels, no words, no extra characters.

The prompt does four jobs. It assigns the inputs, names the important identity details, places the robot in the scene and excludes layout labels. Matching perspective and a contact shadow also gives the subject a physical relationship with the ground.

For your own project, adapt this structure:

Use image 1 for [subject identity], image 2 for [setting or base], and image 3 for [layout or style]. Create [one finished scene]. Place [subject] at [position and relative scale], interacting with [specific surface or object]. Preserve [essential details]. Match [light direction, perspective and material treatment]. Do not copy [unwanted guide text or reference background].

This template is a starting instruction, not the prompt behind a separately demonstrated result. Fill in concrete relationships: “the bottle rests on the marble counter” is more useful than “blend everything naturally.”

For a fixed base, change the first sentence: “Use image 1 as the base; preserve its camera angle, architecture and existing objects.” Then identify the addition and the area it may occupy.

5. Compare the result with each source, not just the overall impression

The same brief had also been attempted with only the robot reference. That version looked coherent: the robot stood on the right and retained its identifying features. However, the model invented a conservatory with a different roof and lighting arrangement.

With only the robot reference, the setting came mainly from the written description. A convincing conservatory was not necessarily the supplied conservatory.

The three-reference version incorporated the provided glass roof structure, warm doorway and walkway more closely, while retaining the robot's asymmetric eyes, antenna, backpack and blue hands. It did not reproduce the layout guide's labels.

Three-reference result: recognizable subject, a setting closer to the source and right-side placement. The left side still contains plants and architectural detail, so it is not a finished blank panel for text.

Both images are 2560 × 1440. This pair illustrates what the additional setting and layout references contributed to one task. Because the prompt also changed, it does not isolate reference count as the only cause or establish a general success rate.

Use this review order:

• Identity: do the features that define the subject survive?
• Setting: are the required architecture, product surfaces or landmarks present?
• Placement: is the subject the intended size and in the intended part of the frame?
• Integration: do the feet or object base meet a plausible surface, with compatible perspective and shadows?
• Delivery: is the crop usable, and is there genuinely readable space for the copy you plan to add?

The final question matters in the example. “Left third clear” kept the robot out of that area, but did not turn a planted conservatory into a flat background. For an advertisement, either request a quieter left-side area or add a designed text panel in a layout editor.

6. Correct one visible problem at a time

Keep the accepted image and the original references. Explain what should change and what should stay fixed. Avoid replacing the whole brief with “make it better.”

These are suggested corrections, not demonstrations that every one has succeeded in this example. If a local correction repeatedly changes good areas, move to a tool with selection, masking or layers. For lettering, finish with the guide to editing text in images online.

Can you do the same thing in other tools?

Dreamina is the reference-led starting route in the workflow above. If you prefer another interface, use the instructions for the model you actually select. In ChatGPT Images, upload the pictures in a conversation, identify their roles and request a single composition. Its official help page also describes edits through chat or a selected area, while warning that changes may extend beyond a selection.

Midjourney's current Edit Model accepts up to four references on versions 8.1 and 8.2. Leonardo's newer Omni Image Guidance supports up to six; older guidance controls work differently. These counts are interface capabilities, not reasons to fill every slot.

For most small composites, the better starting question is whether the tool respects your essential subject and scene, not which one accepts the longest list of files. If the same character must recur across separate pictures, use the character-consistency workflow comparison. Combining references for a still image also differs from combining references for video, where motion and continuity need their own checks.

Save the accepted image at its actual downloaded size, keep the source set and record the prompt that produced it. You are finished when the subject is recognizable, the important setting survives, the elements belong together and the crop serves its intended use. Start with your references in Dreamina, give each image a clear role and judge the result against that brief.