An AI roller coaster selfie needs to keep the rider's reaction in view as the car turns. This Halloween example builds a continuous shot in CapCut with the phone facing the same fictional adult, one arm reaching toward the lens, and the other hand holding the lap bar.
The useful distinction is what appears beyond the camera: in a selfie, the attraction moves behind the rider's shoulders. A view looking down the track answers a different brief. This tutorial focuses on that camera relationship and the checks that made the downloaded take usable.
We tested the web workflow on October 21, 2026. The reference person and attraction were generated in CapCut; this is a fictional scene, with the native AI indicators retained.
Start your AI roller coaster selfie with the holding pose visible
Your reference should show enough of the pose to make the selfie readable. In our source, the whole head and upper torso fit inside a landscape frame. A forearm extends into the lower-left foreground, while the hand on the right of the image wraps around the metal bar. The red seat locates the rider inside the car.
The actual uploaded reference. The unseen phone is the camera making this image; a second phone does not need to appear beside the rider.
We generated this original reference in CapCut Design Studio with GPT Image 2 selected, requesting 16:9 in the prompt. Its native PNG measured 2048 × 1152. To create a similar starting composition, describe a fictional adult seated in a red roller-coaster car, with his whole head and upper torso visible, a forearm extending into the lower-left foreground toward the unseen selfie camera, and the hand on the right of the image gripping a metal lap bar. Specify a rust-orange jacket, a cream shirt, and generic Halloween scenery. Check the generated pose before using it as the video reference.
A tight face crop leaves the holding pose outside the reference. For this task, choose a frame that also lets you inspect the forearm and restraint before generating motion. We did not run a separate face-only reference comparison, so this is a selection criterion rather than a claim about how that alternative would turn out.
Use image-relative directions when describing the hands. “The forearm at the lower-left of the image” is clearer here than assuming which anatomical hand a mirrored selfie shows.
Describe what carries the camera through the curve
Open CapCut's AI video generator and enter Video Studio from the workspace sidebar. For this AI moviemaker task, the reference defines the starting view; the prompt describes how the phone stays with the rider during motion.
1. In the composer, use + → Upload from device and select the reference PNG. Wait for its thumbnail to load in the input before sending. In this run, the filename appeared while the thumbnail was still loading.
2. Choose Video and open the settings beside it. We selected Video clip, Seedance 2.0 Mini, 720p, 16:9, and 10s. The composer mode was Standard. The completion reply later called the model Seedance 2.0 Fast; these names record the differing interface labels rather than establish which backend model generated the clip. Check the options and any credit or trial indicators shown in your own account before submitting.
The settings are shown with the same selections restored after the run. Amber outlines mark Video clip, the model and resolution, 16:9, and 10s.
3. Add the prompt below. The key relationship is that the camera sits at the end of the extended arm and travels with the rider. “Front-facing camera” is wording inside the prompt; we did not use a camera-facing switch in the interface.
Animate the uploaded fictional adult in one continuous 10-second horizontal 16:9 haunted roller-coaster SELFIE. Preserve the same curly-haired man, rust-orange jacket, cream shirt, red seat and metal lap bar.
The phone's FRONT-FACING camera stays at the end of his extended arm on the LEFT of the image and faces back toward his face and upper torso for the whole shot. The phone itself is outside the frame. The selfie forearm enters the lower-left foreground; his hand on the RIGHT of the image stays wrapped around the lap bar. He remains seated and secured. Keep his recognizable face visible as the main subject. The camera travels with his hand and the car; it does not orbit him or turn forward to film the track.
0-3 seconds: already moving, small irregular wheel vibrations, a nervous smile. Amber carnival lights and wooden supports slide behind his shoulders.
3-7 seconds: the car rounds one clear curve. His head and shoulders sway slightly, and the phone tilts with his wrist then recovers, still facing him. A plain sheet-ghost attraction prop passes behind his shoulder. The background moves while his face stays at roughly arm's-length distance.
7-10 seconds: he briefly widens his eyes, then laughs with relief while the ride continues. Keep the lap bar and gripping hand connected; let the face drift slightly rather than freezing it in place.
Casual low-light phone footage with modest motion blur and small exposure changes, enough light to recognize the face. Original generic Halloween scenery, no franchise characters. Avoid a forward-looking track POV, third-person camera, drone shot, cuts, extreme drops, extra hands, identity changes, text, logos or visible second phone.
4. Send with the arrow beside Standard. Our first submission returned Failed: Generate videos in chat while the canvas card still displayed 99%. We returned to the composer and sent the same loaded reference, prompt, and settings again. That retry returned a playable video. The percentage alone did not establish that a finished clip was available.
Read the turn behind the rider
The accepted download measured 1280 × 720, with 241 picture frames at 24 fps, just over ten seconds. The camera kept facing the rider as the background changed from warm interior beams to a gray fence and back to darker scenery. The holding arm stayed in the foreground, and the other hand remained on the bar.
Four frames from the accepted native download, read across each row. The camera keeps facing the rider as the scenery turns behind him.
Use the following moments to judge the relationship in your own take. Times are approximate observations from this export, rather than promises that the prompt will place an event at a precise second.
The prompt's time ranges were a shot plan. The ghost was already visible at the start of the returned take, and the lighting became brighter than the requested dark interior. We accepted the shot for the continuous selfie task. A brief requiring a fully enclosed dark ride or a precisely timed first appearance would need those additional conditions checked before acceptance.
Download the take and follow the hand around the bar
5. Use the download icon beneath Video 1. Review that file at normal speed, then pause before, during, and after the turn. Compare the face with the reference and follow the gripping hand into its sleeve. The arm should still connect to the same body, and the bar should remain in front of the seated rider.
If a future take changes to the track ahead, loses the rider, or breaks the arm-to-body connection, set it aside for this brief. Return to the source composition and the camera relationship before adding more action. We did not need a prompt revision for the accepted result, and we have not tested those changes as repairs.
When deciding how to make video edits around this reaction, keep the turn and recovery together long enough to read the expression. This example verifies one generated shot; it does not cover a measured track layout, a cut to another angle, or stabilization settings. Save the reference and prompt with the downloaded take so you can check the same holding pose when you make the next version.