A gentle-giant creature short needs two things in the same shot: a creature that looks dangerous and an action that is kind. In the first clip I generated, a one-eyed giant handing a teddy bear back to a small girl, the dangerous half was gone within a second. Its bared teeth were covered by 0.75 seconds, and by 4 seconds it was grinning. Nothing in the prompt asked for that. This article follows one AI-generated still through seven clips in CapCut to find what kept the giant frightening while it did something gentle: which words softened its face, which ones changed the speed of its hand, and where music had to carry the rest.
The finished short: one 10-second clip with a music-box lullaby added in the editor. The GIF plays at 6 frames a second and has no sound; the lullaby is in the video file.
Let the houses set the size
The size comes from the frame, so I set it in the still. It came from the AI image generator in Image mode (Seedream 4.5, 16:9, 2K), at 2560×1440:
A cinematic photograph of a quiet village street at dusk after rain. A one-eyed giant, taller than the two-story stone houses beside it, crouches in the middle of the street. It has rough gray-green skin, long cracked fingernails, a heavy brow over one huge yellow eye, and a wide mouth of crooked teeth. At its feet, a little girl about six years old in a red raincoat stands looking up at it. Her bright yellow teddy bear lies in a puddle between them. The giant looks down at her. Low-angle wide shot from street level: the girl is small at the bottom of the frame, and the giant fills the upper two-thirds. Its bald head is level with the rooftops on either side. No text anywhere.
The low camera does the work: the girl stands about 260 pixels tall in the 720-pixel frame, the giant's head touches the top edge, and the stone houses on both sides give it a known size to be compared with. The bear matters as much as the giant. It is the one object the gentle action is about, and its bright yellow makes it easy to follow in every frame.
This was the second try. The first ended with “the giant fills the upper two-thirds, with the rooftops beside its shoulders”, and the picture put a slate roof on top of the giant's head. Saying where its head was relative to the roofs, “Its bald head is level with the rooftops on either side”, kept the roofs on the houses.
The first and second stills, and a close-up of the giant's head in each.
The snarl lasted less than a second
Every clip came from that still in the AI video generator, in Video mode with the still attached and Seedance 2.0 Mini, 16:9 and 720p. The first prompt described only the action:
The giant picks up the yellow teddy bear from the puddle and gives it back to the girl.
The action came back as written: the giant reached down, lifted the bear and held it out to her. Its face changed with it. The teeth bared in the still were covered by 0.75 seconds, the mouth had settled into a closed smile by 2 seconds, and by 4 seconds it was a wide grin. The creature that was frightening in the still ended the clip as a friendly one, and the contrast went with it.
The action-only prompt. The GIF plays at 8 frames a second and has no sound.
What “gently” and “kind” did
The usual advice is to write the gentleness in, so the next two prompts did:
The giant very slowly and gently picks up the yellow teddy bear from the puddle and gives it back to the girl.
The kind, friendly giant smiles warmly, picks up the yellow teddy bear from the puddle and gives it back to the girl.
Neither slowed the hand. I tracked the bear and timed its rise from 30 to 150 pixels above the puddle, in the first frame's 1280-pixel-wide coordinates, so that camera moves don't count. The action-only clip lifted it at 131 pixels a second. “Very slowly and gently” lifted it at 160, and “The kind, friendly giant smiles warmly” at 288, while the camera pushed in to 1.28 times the size of the first frame. Both faces went the way of the first clip, sooner. The “gently” clip showed one fang at 1 second and a wide grin full of teeth by 4. The “kind, friendly” clip was grinning by 0.42 seconds, as asked, and squinting with it.
Write the face into the prompt
If the face changes by default, it has to be written down. The next prompt kept the plain action and added one sentence about the face:
The giant picks up the yellow teddy bear from the puddle and gives it back to the girl. Its face does not change at any moment: its heavy brow stays low over its one unblinking yellow eye, and its crooked pointed teeth stay bared.
The brow stayed low, the teeth stayed showing and no smile appeared in the 5 seconds. The hand did the same kind thing as before, and the bear rose at 120 pixels a second. This is the version where the two halves exist together: the face says danger, the hand says the opposite.
The giant's face in five clips from the same still. Only the bottom row's prompt described the face.
Slow comes from a time slot
Adverbs didn't slow the hand, so the next prompt gave each part of the action a time instead:
For the first two seconds, the giant only looks at the girl. Then, over the last three seconds, it lowers one huge hand, lifts the yellow teddy bear out of the puddle between two fingers and holds it out to her.
The giant didn't wait two seconds: I measured the motion of its arm against the houses, and the arm was moving from 0.75 seconds. The lift still came later than in any of the clips without a time slot: the bear left the puddle at 3.25 seconds and rose at 111 pixels a second. The face, which this prompt didn't mention, blinked and settled into a closed smile within the first second.
In 5 seconds there is little room for a slow reach, so the finished version asked for 10. With the duration set to 10s, the composer's cost line read 50 credits for the request, where 5s had read 25:
For the first three seconds, the giant only stares down at the girl without moving. Then, over the next five seconds, it lowers one huge hand, lifts the yellow teddy bear out of the puddle between two fingers and holds it out to her. In the last two seconds, she takes the bear and hugs it. Its face does not change at any moment: its heavy brow stays low over its one unblinking yellow eye, and its crooked pointed teeth stay bared.
The giant held still for 2.04 seconds of the three it was given. The bear left the puddle at 5.25 seconds and rose at 85 pixels a second, the slowest of the seven clips, and she took it at about 8.75 seconds and hugged it. The face held for all 10 seconds. Of the three slots, the last two landed close to their times, and the first came up about a second short.
The bear's height above the puddle, in girl heights, with the rise speed of each clip in the legend.
The 10-second clip used in the finished short: the face through the clip, and three moments of the action.
Two ten-second takes, two different slips
The 10-second clip changed one thing the still had set: the girl's hood, which is down in the still, came up over her hair within the first second. I sent the same prompt again with one more sentence, “The girl's hood stays down the whole time, and her dark hair stays uncovered.” Her hood stayed down for all 10 seconds. The giant's face didn't hold this time: the snarl eased from about 2 seconds, with fewer teeth showing, and from about 6.5 seconds, as she reached for the bear, the giant smiled. The bear rose at 120 pixels a second.
So the same face sentence held for the full 10 seconds in one take and for about 6.5 in the other. I kept the first take: the hood rises early and small in the frame, while the smile would have undone the contrast in the shot's last seconds. If you run two takes, compare them on both: the giant's mouth late in the clip and anything the still showed on the girl.
Take 1 kept the face and lost the hood; take 2 kept the hood and lost the face.
Let a lullaby do the kind part
The clip came back with its own sound, and that sound took the giant's side. It was quiet for the first two seconds, about −42 dBFS. Then, from about 2.4 to 4.6 seconds while the hand came down, it turned low and loud: 81% of its energy was below 150 Hz, peaking at −22 dBFS. With a face that says danger and a low-pitched sound under the hand, the kindness had to come from somewhere else. In the editor, with the 10-second clip on the main track:
In the export the lullaby runs at about −21 dBFS. The clip's own sound came through at its original level, so its low-pitched peaks reach about the level of the music, and the rest of it sits about 20 dB under the music. The picture is the clip's, frame for frame: 241 frames.
The request behind the finished clip, and the editor steps for the lullaby.
The words to cut from a gentle-giant prompt
These are the words that softened the giant or sped up its hand in these clips, and what I used instead.
Leave the kindness to the action and the music, describe the face you want to keep, and give the reach its own seconds. Then check the clip at about 2, 6 and 9 seconds for the two things that slipped here: the giant's mouth, and anything on the girl that the still showed.
This article doesn't cover dialogue, a story longer than one shot, or footage of a real child. I made two stills and seven clips with Seedream 4.5 and Seedance 2.0 Mini at 16:9, and tried 5s and 10s only.
Configuration and cost. Stills: Image mode, Seedream 4.5, 16:9, 2K, where the composer's cost line read 1 credit for each; two in all. Clips: Video mode, Seedance 2.0 Mini, 16:9, 720p, with the still attached, where the cost line read 25 credits for each of the five 5s clips and 50 credits for each of the two 10s clips. The account's credit details for September 29, 2026 list two charges of 1 credit, five of 25 and two of 50, 227 credits in all, and the balance in the header went from 24,157 to 23,930. No request failed. Nothing done in the editor, including the music and the export, appeared in the credit details. These figures apply to those settings on that date; other models, durations, resolutions and modes are charged differently.
Files. The stills are 2560×1440. The clips are 1280×720 at 24 frames per second, 121 frames for 5s and 241 for 10s, each with its own sound. Each generated still and clip carries a small “Ai” mark in its top-left corner. The export is 1920×1080 at 24 frames per second, 241 frames, and its picture starts 0.083 seconds after its sound. The GIF of the finished short is 480 pixels wide at 6 frames a second, the setting for clips of 10 seconds and longer; the other GIF is 560 pixels wide at 8 frames a second. Prompt text is quoted as sent. The composer's reply to the first 10-second request named the model as Seedance 2.0 Fast; the settings line under the request and the cost line read Seedance 2.0 Mini and 50 credits.
How the face was read. By eye, on close-ups that follow the giant's eye (tracked as a bright yellow disc) every half second, and every sixth or tenth of a second around each change.
How the hand and the bear were measured. Arm: optical flow in a box over the giant's arm, minus the camera motion fitted on the houses; the arm counts as moving from the first moment its motion stays 1 pixel a frame above its level in the first quarter second, for a quarter second. On a test sequence that only zoomed for 11 steps and then moved a patch 3 pixels a frame, it reported the move at step 12. Bear: the yellow shape followed frame by frame with the giant's eye excluded, and its positions mapped back to the first frame's coordinates using the camera motion fitted on the houses. A planted yellow disc rising in a zooming frame was found in 40 of 40 frames, within 0.5 pixels.
How the sound was measured. Level in dBFS over 50-millisecond windows, with the share of energy below 150 Hz; test tones planted at 1.00 and 2.00 seconds were found at 1.00 and 2.00 seconds. In the export, the clip's own sound was lined up by cross-correlation (0.046 seconds) and came out at a gain of 1.00 against the clip; the lullaby's level is what remains after subtracting it.