You want a ghost party with a cup holder, a tambourine player and a waving guest, each keeping its own prop. This AI moviemaker workflow in CapCut builds that scene from three individually generated action clips, so you can replace one failed performer while keeping the other takes.
The characters and stage are original AI-generated demonstration assets. We tested this web workflow on October 22, 2026. The result is a flat illustrated composite with independent character movements. It does not demonstrate a prop handoff, synchronized singing or a moving 3D camera.
Four frames from the same accepted native motion export, read across each row at 0.000, 1.667, 3.333 and 5.033 seconds. The HTML edition shows the movement as a GIF. Native AI indicators remain visible in the composite.
Assign each ghost party role to a separate clip
Give each performer one recognizable item before generating movement. In this example, the blue bow tie identifies the cup holder, the gray fabric identifies the tambourine player, and the purple cone hat identifies the waving guest. Keep the stage empty so that it does not contain a second copy of a performer.
1. In Video Studio, use the Image creation mode to make the separate references. Our CapCut AI video generator session used the Seedream 4.5 selector with 16:9 and 2K selected for these images. Inspect the downloaded files: the full hem, hand and prop need space around them. The first tambourine reference cut into the lower body, so we generated a wider shot before animating it.
The accepted references and empty stage used in this run. This board assembles separate native images for review; it is not a single generated group scene.
The original reference files are the cup holder image, tambourine player image and waving guest image. Their downloaded dimensions were 2560 × 1440. The stage input was a 1280 × 720 JPEG prepared from the empty panel of a CapCut-generated contact sheet.
2. Select a reference card and choose Image to video. Check the settings again after changing modes. In this session, that action reopened Seedance 1.0 mini with Auto as the aspect ratio; we selected 16:9 again for the accepted requests. All three used Video clip, 5s and 720p. The accepted cup take used the Seedance 1.0 mini selector; the tambourine and hat takes used Seedance 2.0 Mini. The cup clip measured 1248 × 704; the other two measured 1280 × 720. These are the observed selectors, not an independent identification of the backend model.
Describe one small action for the selected reference. The hat request asked for a friendly wave and gentle body bob. The tambourine request asked for small wrist shakes and a sway. Keeping each action in its own file makes the later edit a choice about that performer, rather than another request to generate the whole party.
Keep the take with one cup on the chosen side
3. Download and inspect the solo result before arranging the scene. Our first cup take moved the cup to the opposite side of the body. The blue bow tie remained visible, but we wanted the cup to stay on the right side of the image. A later take produced a second cup. Neither was suitable for the intended role, even though both were playable videos.
Read across each row. The rejected take moves its cup across the body between 0.00 and 3.75 seconds; the accepted take retains one cup on the right at the start and 5.00 seconds.
We returned to the reference image and replaced the cup action while retaining the accepted hat and tambourine files. This was the final cup request:
Animate the supplied reference as a single locked-off front view. ONE ghost with ONE orange cup on the right side of the image and its original blue bow tie. The free hand on the left side of the image stays EMPTY and lowered beside the body in the original pose. Both arms and hands stay nearly frozen in their reference positions. Only a slow gentle body sway. Keep the ghost on the ground. Preserve the two black oval eyes without adding pupils or a mouth. One cup total, no extra cups, no new props, no duplicates, no spinning, no turning, no zoom, no cuts. Keep the green background.
The accepted take still has a small body movement. It is not a large toast or a dance routine. Separate generation did not prevent errors inside a solo clip; it let us reject that clip without replacing the other performers.
Remove green without losing the fabric or the eyes
4. Create a project in the CapCut online video editor and upload the stage and accepted solo videos. Put the stage on the main track and move each character video onto its own track above it. Align their starts at zero. We muted the three imported video tracks with their speaker controls for this visual example.
Select one character clip and open Smart tools → Remove background. Enable Chroma key, choose the eyedropper and sample a clear green area beside the character. In our run, a stronger key removed portions of the gray fabric; raising the hat key to 55 also damaged the black eyes. Inspect the character as well as the disappearing background.
The gray clip's accepted key values. Lower intensity preserved more fabric, but left colored edges that needed a separate color adjustment.
5. For these clips, open Basic → Color adjustment → HSL, select the fourth, green swatch, and set its Saturation to −100. We applied this individually to the three character clips. It reduced the green tint in the remaining edges while the orange cup and yellow tambourine stayed visibly colored in the export. Leave the stage's color controls separate.
The green-channel adjustment used on the gray clip. Review the saved result against the party background; these values are specific to the supplied inputs.
The tambourine clip also left a faint rectangular background patch outside the character. We opened Basic → Mask → Rectangle and unlinked the Size fields. The accepted mask used W 400, H 400, Position X 0 / Y 0, Feather 0 and Rotation 0. It retained the tambourine and both moving hands while excluding the distant patch. Check the whole action before choosing Done; the rectangle needs room for the extended hand as well as the body.
Leave room for hands on the shared stage
6. Use Basic → Transform to arrange the performers. The first layout put the tambourine player's extended hand close to the cup holder. We reduced the middle clip's Scale from 80% to 70% and moved it down to keep its hem near the same floor area. The accepted values were:
Shadow stayed at 0 for the three keys, and the clip rotations stayed at 0°. These are settings for our source files. Position your own characters by their visible bodies and full arm movements; their empty green margins may differ.
The accepted layout keeps the solo movements on separate tracks. The middle character is smaller so that its extended hand has more room beside the cup.
If you are learning how to make video edits with several AI performers, check the widest gesture before fixing the spacing. The first frame alone did not show how far the tambourine player's free hand would extend. This scene remains a flat composite; the shadows and expressions were not regenerated together as a single physical scene.
Cover the final picture frame with the background
7. Choose Export → Download. We used 1080p, Recommended quality, 30fps and MP4. The solo videos contained 121 picture frames at 24 fps, lasting about 5.04 seconds. A stage ending at the editor's 00:05:01 left a black background on the final exported picture frame.
We extended the stage to 00:05:02 and exported again. The accepted picture stream measured 1920 × 1080, 152 frames at 30 fps, about 5.07 seconds. Its last frame still contains the stage and all three performers. The audio container runs slightly longer; it is not the picture duration.
On your saved file, follow the orange cup through the body sway, the yellow ring through the wrist movement and the purple hat through the wave. Check the tambourine and free hand against the mask boundary, then inspect the last picture frame. Keep that accepted composite with its three solo inputs so that your next change can replace the chosen performer without rebuilding the party.