Putting two recurring AI characters in one image is harder than generating either character alone. A model can reproduce both faces separately, then blend them, swap their features or attach the right face to the wrong body when they share a scene.
The fix is not a longer description. You need to make identity assignment easy: give each character a separate reference, a fixed position and a distinct visual role. Then build the interaction in stages instead of asking for a complex final scene in one generation.
Why two-character images go wrong
A two-character prompt asks the model to preserve two identities, place each identity on the correct body, maintain clothing and scale, handle overlaps, and render the setting. When the people have similar hair, clothing or facial structure, the assignment becomes even less clear.
Names such as "Maya" and "Leah" may help you organise the prompt, but they do not give the model much visual information. Start by making each person unmistakable.
1. Prepare a separate reference set for each character
Do not combine every source photo into one unlabeled collage. Keep a small, clean set for Character A and another for Character B.
For a waist-up scene, a clear front view and three-quarter view of each person are usually more useful than distant full-body shots. For a standing scene, include a full-body reference that establishes height, build and clothing proportions. The references should show the same intended age, hairstyle and defining features.
If you are still building those source images, create an AI character reference sheet first. Use the fewest references that clearly establish identity. Extra images with conflicting hair, makeup or lighting can create ambiguity. Our guide to choosing the number of reference images explains that trade-off.
Only use images you own or have permission to use.
2. Give the characters distinct visual anchors
Two characters can share a visual style without looking interchangeable. Give each person two or three persistent anchors that remain visible at the intended camera distance, such as:
- different hair silhouettes
- contrasting outfit colours
- a stable accessory
- noticeably different height or build
- different sleeve lengths or collar shapes
Avoid nearly identical neutral outfits in the first attempt. You can simplify the styling once you have a reliable joint reference.
Describe visible traits, not vague personality labels. "The confident founder" may affect expression, but "the woman on the left with a short black bob and green blazer" gives the model a concrete identity and location.
3. Block the composition before adding interaction
Begin with both faces visible and no overlapping bodies. A waist-up two-shot with the characters side by side is a good starting point.
Keep the left and right positions, camera distance, outfits, relative height and facing directions fixed. Use a plain or lightly detailed background at first. A crowded location makes it harder to tell whether a bad result came from identity, pose or composition.
Once the neutral two-shot works, add the real environment through an edit. Both Google's image-generation guide and OpenAI's image-generation documentation describe iterative, multi-turn image editing. That workflow is better suited to careful refinement than repeatedly rebuilding the whole scene.
4. Write the prompt in character blocks
Do not mix both people into one long sentence. Separate identity, placement and action. For example:
Scene: Create a realistic waist-up editorial photo of two friends in a bright café.
Character A, left: Use reference set A for identity. Woman with a short black bob, oval face, green blazer and small silver hoop earrings. She remains on the left.
Character B, right: Use reference set B for identity. Man with curly brown hair, square jaw, tan overshirt and dark blue T-shirt. He remains on the right.
Composition: Both faces fully visible with equal visual importance and natural space between them. Character A holds a coffee cup. Character B looks toward Character A. Eye-level camera, soft window light and realistic skin texture. Do not blend their facial features or swap their hairstyles, clothing or positions.
Negative instructions may help, but the positive assignment matters more. "Character A remains on the left" is clearer than a long list of unwanted results.
Keep the labels and positions stable during revisions. Moving Character A from left to right removes one of the strongest cues connecting identity to place.
5. Add interaction one step at a time
Physical interaction raises the difficulty because hands, arms and faces can overlap. Progress from low contact to high contact:
- Stand side by side with space between them.
- Turn their heads toward each other.
- Add a shared prop or simple handoff.
- Place them at a table or on a sofa.
- Attempt an embrace, dance pose or other close contact.
Save each successful stage. If the next edit causes identity drift, return to the last clean image instead of repairing a heavily distorted result.
For complex poses, create the body layout first, then correct each face separately. If hands become the main problem, use the local-edit approach in our guide to fixing AI-generated hands.
6. Repair one character without regenerating the other
If one face is correct and the other is not, avoid a full-frame regeneration. Mask or select only the drifting face and its immediate hairline. Ask the editor to restore that character from the correct reference while preserving everything outside the selected area.
OpenAI's documentation confirms that masks can indicate which area should be edited, while noting that a mask is guidance rather than a perfectly hard boundary. Leave enough room around the face for natural blending, but do not include the other person.
For advanced local workflows, the Diffusers IP-Adapter documentation shows two face references assigned to two spatial masks. Regional conditioning can help when a general-purpose editor keeps mixing identities.
After every repair, check for swapped hair or accessories, changed age, incorrect body shape, outfit details copied from the other person, mismatched eye direction and altered scale. Fix one problem per edit. Asking for a new face, pose, outfit and background together makes it harder to preserve what already works.
What to do when identities still swap
If the faces keep changing places, return to a front-facing two-shot and strengthen the spatial language: "Character A occupies the left third; Character B occupies the right third." Reinforce the distinction with outfit colour and hairstyle.
If both faces merge into a generic average, use clearer crops, reduce the number of inputs and increase the physical separation between subjects.
If one character becomes a background extra, request equal visual weight, similar face size and both sets of eyes visible. Avoid "main character" and "companion" unless you want unequal framing.
If the composition works but likeness does not, build a rough composite. Generate each person at the required angle, place both cutouts on a canvas, then use image editing to unify the lighting, edges and background. This supplies explicit geometry instead of asking the model to invent it.
Turn one successful image into a duo reference
Once both identities are correct in the same frame, save that image as a joint anchor. Use it alongside the individual reference sheets for later scenes. The duo image establishes relative height, scale and separation, while the individual references retain better facial detail.
Do not rely on the joint anchor alone for every pose. Keep the original character references and use a multiple-pose character workflow when the camera angle changes substantially.
The reliable sequence is simple: establish each identity, create a low-contact two-shot, lock positions, add interaction, then repair locally. That gives the model fewer decisions at each stage and gives you a clean checkpoint whenever a revision goes wrong.
Explore AI character generator for the workflow, then create your character in Rasgo.
Create your next Rasgo visual
Turn the ideas from this guide into generated images, creator content, product shots, or video-ready concepts.