When you make two social posts interact, the viewer needs to see that the characters belong to separate photographs. This CapCut example creates a Halloween image in which a ghost reaches across a feed interface and lifts a fictional adult by her jacket, while the dark hallway and daytime sidewalk stay apart.
The finished image is one 16:9 graphic. Its two posts are drawn inside the image; they are not separate live posts that interact on a social platform. We tested the web workflow on October 21, 2026, using an original AI-generated adult and a generic sheet ghost. The native AI indicator remains in the result.
Leave a visible band between the two social posts
The middle interface gives you a clear place to judge the illusion. In the accepted image, the hallway ends at a straight horizontal edge. A charcoal band follows, then the bright sidewalk begins at a second edge. The woman's shoulders cover part of that band, while the outline icons remain visible around her.
The accepted native PNG. The hand grips gathered jacket fabric near the upper edge of the interface, and the woman's upper body crosses the band while her legs remain over the lower photograph.
You can use the two photo edges as reference lines. A wall or tree continuing through the band would weaken the separation. Here, only the foreground characters bridge the gap. The lit pumpkins stay above it, and the white fence stays below it.
This example focuses on that single-image relationship. It does not animate the characters, create a working social interface, or build a swipeable carousel.
Generate the full composition with both worlds specified
1. Open CapCut AI Design and enter Design Studio. In the composer, open Select model, Auto and choose GPT Image 2. This was the selected model label for our two submissions; we did not independently identify the backend model.
The actual model selector during this run. The amber outline marks the selected GPT Image 2 row. The selection menu is separate from the composition instructions.
2. Enter the prompt below in the composer. It requests the wide composition at generation time. The separate upper-world and lower-world descriptions establish which background stops at each photo edge; the depth paragraph identifies what passes in front of the interface.
Create one original photorealistic Halloween cross-post interaction image in 16:9 landscape, 2048 by 1152 composition. This is a single flat graphic showing TWO stacked fictional social-media photo posts, not a phone or device, with no real brands or logos. The wide rectangular photos fill the canvas width. Top photo about 42 percent, middle charcoal interface band about 16 percent, lower photo about 42 percent. Straight horizontal photo edges. The middle band remains visibly intact on the left and right of the characters and contains small clean white outline heart, comment and bookmark icons, minimal fictional account text only.
UPPER WORLD: A friendly eerie original sheet ghost with two dark eye holes, head and torso clearly visible in a dark blue abandoned wooden hallway with two small warm pumpkins. The ghost is near the center-left. One long fabric-covered arm grows continuously from its visible shoulder, reaches down out of the upper photograph, crosses IN FRONT OF the middle interface band and into the lower post. Show the continuous shoulder-to-elbow-to-wrist path. One hand only at the end of that arm. Its fingers visibly pinch a gathered fold of the LOWER person's orange jacket at the upper back/shoulder, not skin, neck or hair.
LOWER WORLD: One fictional AI-generated adult woman in her late twenties, short dark curly hair, plain orange jacket, cream shirt, dark blue trousers and white sneakers. She is in a bright daytime neighborhood sidewalk photo. Her body is pulled slightly upward by the gathered jacket fabric, both feet visible above the sidewalk, knees bent and arms open for balance. Her surprised face and shoulders overlap the middle interface band while her legs remain inside the lower daytime photo. Keep her face and the ghost's face separate and readable.
DEPTH: The ghost's arm and the woman's upper body occlude a small central part of the interface. Add subtle contact shadows onto the flat middle band. Leave recognizable interface icons visible on both sides of the contact. The dark upper hallway must STOP at the upper photo boundary. The bright lower sidewalk must STOP at the lower photo boundary. The two backgrounds remain separate; only the connected ghost arm and lifted person's upper body cross the band. No blended landscape, no doorway connecting the scenes, no hole or shattered glass. No extra hands, arms, people or text headlines. Playful Halloween comedy, no injury, blood or gore. Generate the complete wide composition as one image.
3. Send the message and wait for the generated image. In our run, the home composer opened a design project and started the request there. The first image already showed a connected arm, gathered fabric, and separated backgrounds. We generated the whole graphic together; we did not test a second method made from separately assembled layers.
The requested 42/16/42 proportions were composition targets. The final image gave the interface roughly 28 percent of the height, leaving a shorter lower photograph. We kept that result because both feet remained visible and the contact was readable. For a brief with fixed photo heights, those dimensions would need a separate acceptance check.
Follow the arm into the gathered jacket
Your review can start at the ghost's shoulder and continue along the sleeve to the hand. In this result, the fabric-covered limb stays connected, and the fingers meet the peak of the orange jacket fold. The woman's neck and hair are outside the grip. The stretched jacket leads into her lifted torso; both shoes sit above the sidewalk.
Four inspection views from the same accepted native still, read across each row. Labels and photo-edge highlights were added locally. The HTML edition shows these views as a review GIF, not as character motion.
The hand sits near the top of the charcoal band because the jacket has been pulled upward. It does not have to remain inside the lower photo rectangle to read as contact with the lower character. What matters here is the visible connection from hand to gathered cloth to the same person's body.
These checks describe the result we accepted. We did not encounter a detached hand or an extra arm in these two outputs, so this is not a tested repair for those faults. If your own image lacks a readable connection or merges the backgrounds, it does not meet this brief even if the characters look appealing.
Clean the interface text before downloading the selected image
The first image added a small decorative symbol after the caption and omitted a separate lower-post account header. Its spatial relationship already worked. We selected that image and used a follow-up request to simplify the caption and add a fictional lower-post name.
4. With the first image selected on the canvas, send the cleanup request below. We kept the same model selection for this submission.
Edit the selected cross-post image, retaining its 16:9 landscape composition. Preserve the two separate photo backgrounds, ghost pose, continuous attached arm, gathered orange jacket grip, woman, lifted feet, straight photo boundaries, middle charcoal interface band and visible outline icons. Make only this interface-text cleanup: remove the small ghost emoji after the caption and replace the caption with the plain words Just hanging around. Keep the fictional friendlyghost name. Add a small plain lower-post profile header reading neighborhoodwalk near the bottom-left of the charcoal band, above the lower photograph. No emojis, real brands, logos or extra text. Keep the characters and the contact unchanged.
The chat briefly showed The network connection was lost. We reloaded the same project to inspect its state before retrying. The completed image was then available in the conversation, so no repeat submission was needed. We applied that generated design to the canvas. The added header was readable, and the connected grip remained intact. The image was regenerated, so the request to preserve the composition should not be read as a claim that other pixels were locked.
5. Open Download and check the export list. We selected only halloween_crosspost_interaction_edited.png, leaving the earlier version unchecked, then used PNG at 1× (2048 × 1152px).
The actual export panel. Amber outlines mark the selected result, PNG size, and Download button. Checking the filename prevents exporting the earlier caption version by mistake.
Our downloaded PNG measured 2048 × 1152. Open the saved file and read the two fictional account names, then trace the arm once more at full size. Keep that accepted PNG with its prompts; use it as the reference for any later seasonal variation before adding movement or changing the characters.