To build a suspenseful opening scene, limit what the audience can see, delay the reveal, and let a sound cue arrive before the threat. This CapCut example starts with one AI-generated still of a woman facing a dark aquarium window. Across five test clips, explicit timing in the prompt delayed the creature more effectively than vague hints. A rumble added in the editor then led the eye's reveal by about a second. The workflow focuses on a ten-second AI opening.
The finished opening combines one 10-second clip with a rumble added in the editor. The preview shown here has no sound; the sound timing described in the text refers to the exported video.
Ask for the creature and it arrives in the first second
The still that every clip starts from was built for a restricted view: her back to the camera, a window of black water, and one small circle of light that is all the audience can see into. I made it with the CapCut Seedream 4.5 image generator in Image mode (Seedream 4.5, 16:9, 2K), at 2560×1440:
These model names, settings, and menu labels describe the workspace used on September 29, 2026. Check the options and credit estimate in your own workspace before generating; availability and costs can differ.
A cinematic photograph inside an abandoned aquarium at night. A woman in her thirties in a bright yellow raincoat stands with her back to the camera on the left side of the frame, holding a flashlight. In front of her is a huge dark viewing window, taller than a house, into a flooded tank. The water behind the glass is murky and almost black, with fine particles floating in it, and her flashlight beam makes a small pale circle on the glass. Puddles on the tiled floor, peeling paint, a broken railing. Wide shot from behind her shoulder. No text anywhere.
Then I attached it in the CapCut Seedance 2.0 Mini video generator, in Video mode with Seedance 2.0 Mini, 16:9, 5s and 720p, and asked for the scene the way a first draft usually puts it:
A giant deep-sea creature rises out of the dark water and swims right up to the glass in front of her. The camera slowly pushes in over her shoulder.
In the test record, brightened frames show a dark back low in the tank by 1.0 second. By 3 seconds it takes up most of the window, and at 4.5 seconds its face is against the glass. The creature occupies about four of the five seconds, leaving little time to build anticipation. The restricted viewpoint held: the camera stayed behind her shoulder.
The first prompt. The GIF plays at 8 frames a second and has no sound.
Hints turn into a silhouette
The usual advice is to suggest the threat rather than show it, so the second prompt asked only for signs of it:
She slowly moves her flashlight beam across the dark glass. The floating particles behind the glass begin to drift sideways, and the water darkens as if something very large is passing behind it, but nothing comes into the light. The camera slowly pushes in over her shoulder.
The circle of light stayed clear for most of the clip, as asked. The hint didn't stay a hint. By 1.0 second a dark back is visible below the light, by 3 seconds it stretches across the lower window, and at 4.5 seconds the beam lands on it. “Something very large is passing behind it” came back as the creature itself, drawn just outside the light.
Three prompts from the same still, at 1.0, 3.0 and 4.5 seconds. All nine frames are brightened by the same amount so the dark shapes print.
Put the reveal on a clock
The third prompt said what should be on screen, and for how long:
For the first four seconds, only the dark water, the floating particles and her flashlight beam are seen. Then, in the last second, a huge eye, taller than she is, opens right behind the glass in the circle of her flashlight. The camera slowly pushes in over her shoulder.
In this result, the brightened frames from 0 to 94 show water, particles, and the beam. At frame 95, 3.96 seconds, an eyelid parts inside the circle of light, and by frame 97, 4.04 seconds, the eye is open. The reveal began about one frame before the requested four-second boundary. Each prompt was generated once, so this is one observed result rather than a guarantee of frame-accurate timing.
The ten-second version also delayed the reveal. I set the duration to 10s and changed the timing instructions:
For the first eight seconds, only the dark water, the floating particles and her flashlight beam are seen, while the beam moves slowly across the glass. Then, in the last two seconds, a huge eye, taller than she is, opens right behind the glass in the circle of her flashlight. The camera slowly pushes in over her shoulder.
The tank stayed empty through frame 206, 8.58 seconds. An eyelid appears at frame 207, 8.63 seconds, and the eye is open at frame 209, 8.71 seconds. The prompt allowed the reveal from 8 seconds, so it began about 0.6 seconds into that window. Use the wording to guide the sequence, then check the generated timing before editing the sound.
The frames on either side of the reveal in the 5-second clip, the 10-second clip, and the chained clip described next.
A ten-second opening in one request or two
The alternative to a 10-second request is two 5-second clips joined end to end. I saved the silhouette clip's last frame, attached it as the start of a new clip, and used the four-seconds-then-an-eye prompt with Start on the first image added at the beginning. The join looked continuous, but two details weakened the suspense. The hood rose within the first second of the new clip despite being down in the reference. The creature's back was already visible in the first clip and remained visible in the second, so the final eye appeared on a threat the audience had already seen.
The single 10-second result kept the hood down and delayed the creature until frame 207. In this test configuration, its displayed cost was 50 credits, equal to two 5-second requests at 25 credits each. For this short scene, it gave the cleaner result. A longer sequence may still need separate shots; when joining them, check that the first clip preserves the intended concealment and that the next one retains visible character details.
The joint in the two-clip version, and the single 10-second clip at 9.0 seconds.
The generated sound waits for the picture
Every clip came back with its own sound: the hum of an empty room, and a louder, deep sound when the creature showed. I measured the level of each clip in 50-millisecond steps. In the two clips that kept the tank empty, the sound stayed at the level of the room, around −42 dBFS, until the eye. It jumped by more than 20 dB at 4.05 seconds in the 5-second clip, where the eye appeared at 3.96, and at 8.75 seconds in the 10-second clip, where the lid appeared at 8.63. In the first clip, the level began rising under a second in, together with the creature coming into the tank.
In these test clips, the generated sound followed the visual reveal. An earlier sound cue would give the audience time to anticipate it. Adding that cue in the editor makes its timing easier to control and lets you adjust it to the actual reveal frame.
A rumble one second ahead of the eye
In the tested web editor, put the 10-second clip on the main track. Open Audio, select Sound effects, and search for rumble. Preview a low, steady effect, then add it on a separate audio track so it begins about one second before the eye becomes visible. Trim its end to the end of the picture. Select the effect and use its volume and fade controls to keep it beneath the reveal sound. In this test, the Basic panel's Volume was set to −10 dB with a 1-second fade-in; that is a gain setting for this effect, not a target output level in dBFS.
In the measured export, the rumble became audible at 7.70 seconds and the eye was fully open at 8.71 seconds, giving roughly a one-second lead. The rumble before the reveal measured about −22 dBFS, while the generated boom peaked at −9 dBFS. These readings describe this mix; they use different level measures and should not be treated as a universal 13 dB mixing rule. Listen at a comfortable playback level, keep the anticipation cue audible, and lower the reveal if its jump is too abrupt. Export the completed sequence at the frame rate used for editing; this test used 1080p, 24 fps, and MP4.
Top: the sound each clip came back with, with the reveal marked. Bottom: the finished opening against the 10-second clip's own sound.
The request behind the finished clip, and the editor steps for the rumble.
A ten-second timing sheet
Use this schedule as a starting point for a ten-second scene. The prompt targets an eye reveal after eight seconds; in this result, the first eyelid appeared around 8.6 seconds. Align the added sound to the reveal you actually receive.
Adapt two details for your own scene: what the light reveals and how long the audience waits. Check the result at normal playback brightness first, then raise the brightness to look for a silhouette at the edge of the beam. A shape that appears early can weaken the reveal even when the prompt keeps it outside the light.
This example tests one still and five clips with Seedream 4.5 and Seedance 2.0 Mini at 16:9, using 5- and 10-second requests. Each request was generated once; the comparison shows how these particular results behaved. For an audience sensitive to sudden reveals, soften the boom or end before the eye appears.
Test settings and credits. On September 29, 2026, this workspace showed 1 credit for the Seedream 4.5 image at 16:9 and 2K. Seedance 2.0 Mini requests at 16:9 and 720p with one reference image showed 25 credits for 5 seconds and 50 for 10 seconds. One image, four 5-second clips, and one 10-second clip totaled 151 credits in this session. These are dated observations for this account and configuration; use the current estimate in your own workspace when budgeting.
Source and export files. The still is 2560×1440. The clips are 1280×720 at 24 fps, with 121 frames for the 5-second results and 241 for the 10-second result, each with audio. The finished 1080p export contains 241 video frames. In this export, the video stream starts 0.083 seconds after the audio stream, so exported picture timestamps include that offset. Prompt text is quoted as submitted.
Measurement notes. Visual onset means the first brightened frame where the creature can be distinguished from the tank's plants. At normal brightness, the shapes in the first two clips were harder to see before about 1.5 seconds. Sound levels were measured in 50-millisecond windows; the added rumble was measured after aligning and subtracting the generated clip's audio from the export. These methods describe this test rather than settings a reader must reproduce.