In a lineup reveal the viewer is waiting for one thing per player: the moment he stops posing and turns into the camera. Everything else in the template — the lighting, the shirt, the caption — exists so that the moment lands in the same place, eleven times running, and the whole thing reads as one show rather than eleven clips.
So that is what I measured. One prompt skeleton, four invented players, four stills and four short intro clips, and a check on every property the eleven-clip version would need to repeat.
The skeleton
Identical for every player except the two bracketed parts. A cinematic portrait of a football player in a dark studio, lit by a single hard key light from above left and a cold blue rim light from behind. [One sentence of appearance.] He stands facing the camera with his arms folded, chest up, chin level. He wears a plain deep green football shirt with the number [N] in white on the chest and a small plain shield badge with no lettering. The background is black with faint drifting haze. Framed from mid-thigh up, centred. No text, no logos, no real club or league marks anywhere in the frame.
Yellow ring: a lamp nobody asked to see. Red box: the shirt number, which came back right in all four.
What the skeleton held, and what it did not
Two of these matter for the finished film. The badge first: asked for a plain shield with no lettering, one came back plain, two came back with an ornamental pattern, and one came back with pseudo-lettering across the top. It is not a real club's name, and it does not need to be one to be a problem — a crest with writing on it reads as a club, and you are the one who has to check what it says. Look at every badge at full size before a clip goes into the set.
The lamp is the other, and it is the same mechanism as a camera sentence putting a tripod in the room: em>lit by a single hard key light from above left/em> was read as a description of the scene, so the scene got a lamp. It appeared in all four, in the same upper-left position, which at least makes it consistent. If you do not want it, describe what the light does to him rather than where it stands.
The motion sentence, run four times
Appended to each still for image-to-video. Slow steady push-in toward his face over the whole clip. At the end he unfolds his arms and looks straight into the lens. The haze drifts slowly. The light does not change.
The four clips in sync, top row No. 9 and No. 1, bottom row No. 5 and No. 10. Watch the top left and bottom right race ahead.
How far each clip has pushed in, relative to its own first frame, measured by matching image features frame to frame. Dots mark the moment the arms open.
The same sentence produced two different camera moves. No. 9 and No. 10 finished at ×1.96 and ×1.79 — head and shoulders by the end. No. 1 and No. 5 finished at ×1.26 and ×1.23, still close to full figure. Put those four end frames side by side and they look like two different shows.
And em>at the end/em> was not where the arms came down. In all four the reveal happened in the first half of a five-second clip, between 1.50 and 2.50 seconds, after which the player stands still while the camera keeps going. The model treated the order of events in the sentence as an order, not a schedule.
Read down a column to compare players at the same moment. At 1.5 seconds No. 9 is already open; No. 1 is still folded at 2.0.
Cut on the reveal, not on the clock
If the prompt cannot put the reveal at a fixed time, the cut can. Find the moment each player's arms open — it is the largest movement in the clip once the camera move is subtracted, and it is visible by eye on a frame strip — and take a fixed window around it: three quarters of a second before, three quarters after.
Cut on the clock and you inherit the full spread of the two camera moves. Cut on the reveal and every slot starts at nearly the same framing, because you are taking each clip from before its camera had time to diverge. And every player opens his arms at exactly the same point in his slot, which is the one thing the viewer is watching for.
The four clips cut on their own reveals and joined: six seconds, four slots of 1.5 seconds, arms opening at 0.75 seconds into each.
First frame, reveal and last frame of each slot, pulled from the joined file by frame number.
For a full side of eleven, that is 16.5 seconds of intros before any title or team card. Generating each clip at five seconds and keeping one and a half means you are paying for the camera move you are going to throw away; that is the price of a template that holds.
Where the caption goes
The conventional place for a name and position is a lower third. Half a second after the reveal — which is when a caption would be on screen — that is exactly where the arms go.
Percentage is how much of each box an arm, the lamp or haze occupies. The upper right is the only box empty in every clip.
So the caption template has three fields — surname, position, number — set in one fixed style in the upper right, entering on the reveal frame. Because the reveal is at the same point in every slot, a single caption timing works for the whole side.
What to do with this
Where this stops
Four players, not eleven, one run each. The spread in camera moves is a spread across four draws, so a full side will show its own range; the method of cutting on the reveal does not depend on how wide it is. The caption finding is specific to this pose — arms folded, then opened wide — and a different reveal action will put the arms somewhere else, so measure the empty corner for your own pose once before building the caption.
Not tested here: cutting to music, transitions between slots, and 4K export. The joined sequence uses straight cuts at the slot boundaries, and at 1248×704 per clip that is the resolution the video model returned.
The part of the template you reuse is a step, not a sentence
It is natural to think of a template as text: the prompt skeleton, the caption style, a duration table. Those are all here, and they are the parts that held — framing, number, caption position. The parts that did not hold were the ones that depend on time, and no wording fixed them. What makes eleven clips cut together as one show is a single repeated action at the editing stage: open the clip, find where he opens his arms, keep three quarters of a second each side. That step is the template. The text only gets you clips worth cutting.
Configuration and cost. Four stills: Image, Seedream 4.5, 16:9, 2K, Standard — the panel declared 1 credit each. Four clips: Video, image-to-video from each still, Seedance 1.0 mini, 16:9, 5 s, 720p, Standard — the panel declared 5 credits each. Total 24 credits, 2026-09-24. Credit costs are tied to that configuration on that date; the same length and resolution on Seedance 2.0 Mini declared 25 credits in earlier testing.
Files. Stills 2560×1440. Clips 1248×704, 24 fps, 121 frames each — not the 1280×720 the 720p label suggests. The joined sequence is 144 frames at 24 fps, four slots of 36 frames. Prompt text is quoted as sent.
How this was measured. Push-in is the scale of a similarity transform fitted between each frame and the clip's first frame by matching image features; on the first frame magnified by known factors of 1.00, 1.05, 1.10 and 1.20 it returned 1.0000, 1.0500, 1.0999 and 1.2002. The reveal is the frame of largest movement in the torso region after the camera move has been fitted and removed, and was confirmed on frame strips for all four clips. Face position is the centroid of the largest skin-colored region in the upper half of the frame; shifting a still by 100 and 300 pixels moved it by exactly 100 and 300. Caption-box occupancy counts pixels that are edges or bright; a box on an outstretched forearm read 40.8% and an empty black box read 0.0%. Every box was drawn back onto the frame it measured and checked by eye.
One judgement of mine was wrong and the measurement caught it: on the contact sheet two of the four players looked noticeably off-center, and the face-position measure put all four within about two percent of the middle. Drawing the centroids back onto the images confirmed the measure, not the impression.
One verification error was caught. The first strip made to check the joined sequence showed the next player in each “last frame” position, which looked like a timing fault in the join. It was not: frames had been pulled by seeking to a time, and on the joined file that seek snapped forward to the next slot's first frame. Counting frames instead showed four slots of exactly 36 frames with the players in the right order at every boundary, and the strip in this article was rebuilt by frame number.