One person stands sharp and still while the street smears past them. It is a five-second AI video effect built from a single photo, and it fails in a way that is easy to see and hard to describe: the whole frame drifts, the subject softens along with everyone else, and what should have been two speeds becomes one. This article builds the effect at 16:9 in CapCut from a generated street photo, and settles the two questions that decide it, which photo can carry it and how the camera instruction has to be worded, by measuring six clips rather than by looking at them.
The photo has to have two regions in it already
The effect is a contrast between two areas of one frame: a subject that does not move, and a field that does. Nothing in the request can create that division if the picture does not already contain it, so the first decision is about the photograph, and it is a decision you can check by eye before spending anything.
Two street photos were generated for this article at 2560 by 1440 through the CapCut AI image generator flow in Video Studio's Image mode, both with the same man, the same coat, the same wet pavement at dusk. In the first he stands in a clear pocket of pavement with the walkers several meters behind him. In the second the crowd presses in at his own distance from the camera, shoulders overlapping his on both sides, one person crossing in front of his coat.
The same man in two photos. The red box is the region that has to hold still, the blue is the region that has to move. In the second photo the two overlap, which is the whole problem.
Those two boxes are also the measuring instrument for the rest of this article. Everything below reads the average pixel change from one frame to the next inside the red box and inside the blue band, and compares them.
"The camera does not move" did not hold the camera
Each clip was made by attaching the photo to the home composer and sending it through the AI video generator flow at 16:9, five seconds, 720p. The first request used the wording most people would write, a description of the look rather than an instruction about the frame: cinematic motion blur, a street full of movement, a long-exposure feel, with the man in the middle of it.
Cinematic motion blur effect. The street is full of movement and the shot feels like a long exposure, with the man in the middle of it all.
The clip that came back is closer at the end than at the start. Matching the man's face from the first frame against every later frame, at the same settings and the same still, his face ends 28 percent larger, growing steadily the whole way through. The crowd he was supposed to be standing against thins out of frame as the shot tightens.
A second wording named the camera directly and still drifted, so the third attempt stopped describing and started listing. It names the shot type, enumerates the moves that are excluded, and states the condition the last frame has to satisfy.
Locked-off tripod shot. The frame does not change at any point: no zoom, no push in, no dolly, no pan and no crop, and the last frame has the same framing as the first frame. The man in the foreground holds one pose for the whole clip, feet planted on the same paving stone, head and shoulders still, looking at the camera, sharp and in focus. The people behind him keep walking quickly across the frame in the direction they are already facing, and they move fast enough that their bodies stretch into long horizontal streaks. The storefronts, the street lights and their reflections on the wet pavement stay exactly where they are.
Face size across the take, measured by matching the first frame's face against every later frame. Same photo, same model, same settings; the wording is the only difference.
On the locked wording the face measures 1.00 for the whole take, apart from three single frames where the match dropped out. The framing at five seconds is the framing at zero. That is the shot the effect needs, because a subject that grows is a subject that is moving, whatever the prompt said about standing still.
One ratio says whether the two regions separated
Holding the frame is necessary and not sufficient. The result also has to have two different amounts of movement in it, and that is a division rather than a judgment: the average per-frame pixel change in the crowd band, divided by the same figure inside the subject box.
The five clips that were measured. The three near 3.7 are the ones that read as the effect; the two below 2.8 are the ones that do not.
The vague wording scored 2.76 and the locked wording 3.50 on the same photo. Running the locked wording through a second model scored 3.71. The gap between the high and low numbers is not subtle on screen: at 3.5 the man reads as a fixed point and the pavement behind him reads as traffic, and at 2.76 the pushing frame drags his edges along with everyone else's.
A second measurement separates the two regions a different way. Fine detail in the subject box, scored by the variance of a Laplacian, averages 235 on the locked take against 30 in the crowd band, so the man carries roughly eight times the detail of the people behind him in the same frame. On the drifting take the same subject figure falls to 151.
The finished clip, 1280 by 720, 24 frames per second, 5.04 seconds, as generated from the locked wording. Shown here as frames in sequence; the full video and a GIF version are supplied alongside this document.
The clip second by second, and three crops at full size. His face is as sharp in the last second as in the first; the walkers behind him are not.
When the crowd walks through the subject
The second photo, the one where the crowd stands at the man's own distance, was run through the same locked wording and the same model. The frame held, at 1.00 for the whole take, so the wording did its job. The separation still fell to 2.65.
The crowded photo animated with the same wording. The red box is the same region as before; twice in five seconds somebody walks through it.
The number that moved is the one inside the box. Per-frame change in the subject region measures 9.82 here against 3.88 on the first photo, because a passer-by crossing in front of the man is movement at the subject's own address. No wording fixes that, and neither does a better model: the box has to be quiet for the ratio to be large, and it can only be quiet if the photograph left a gap around him.
Speeding it up in the editor moves both numbers together
The clip was then uploaded to the CapCut web editor, selected on the timeline, and set to 2x in the Speed panel on the right rail, which shortened it from 5.04 seconds to 2.53.
The Speed panel with the clip selected. Setting the field to 2 turns the five-second take into a two-and-a-half-second one.
On the export, the crowd band rose from 13.55 to 17.62 and the subject box from 3.88 to 4.77, which leaves the ratio at 3.69 against 3.50. Speed changed how fast the crowd crosses, not how much the crowd smears: fine detail in the crowd band reads 29.9 before and 28.7 after, the same band of numbers, because the blur inside each frame was fixed when the clip was generated. Use the speed change to fit a cut, and expect nothing from it in the way of separation.
What this AI video effect cost, and what the model list changes
Every request came from the Video Studio home composer, which prints a cost line before it starts.
The two cost lines this article was charged, each with the configuration it belongs to.
The model is chosen from Model and quality inside the composer's settings popup, and the list carries a one-line description for each entry.
The model list as it appeared in this session. The entry that is checked is the one the cost line above belongs to.
Two of those entries were used here and behaved the same way on this shot: with the locked wording, Seedance 2.0 Mini held the frame at 1.00 and scored 3.50, and Seedance 2.0 Fast held it at 0.96 and scored 3.71, both at 25 credits. The Seedance 2.5 request printed 185 credits, returned three cards reading "Couldn't load content", and the balance in the header, which had fallen by 185 while it ran, was back to its earlier figure afterwards. These figures are for these configurations on 9 September 2026, read from the composer's own cost indicator; other durations, models, resolutions and modes are charged differently, and a credit-discount banner for two of the models was showing on the home page during the session. Of the ten requests, four clips and two photos are in the article; the first three clips were sent before the aspect chip was set to 16:9 and are not used above.
Which photo you start with
Both of the numbers that matter trace back to the picture. The wording decides whether the frame holds, and a locked-off instruction that lists the excluded moves held it here where a description of the look did not. The photograph decides whether there is anything to separate, and no amount of wording moved that: the same instruction scored 3.50 on the picture with a gap around the subject and 2.65 on the picture without one.
So the work happens before the first request. Look at the photo and find the subject's outline; if you can trace it without crossing anybody else, the shot can carry the effect, and if a shoulder or a bag or a coat cuts across it, choose a different frame or step the subject forward and take it again. Then write the camera instruction as a shot, not as a mood, and measure the export rather than trusting it.
One boundary belongs with this particular effect. A real street photograph is full of people who did not agree to appear in anything, and this method makes them the moving half of the picture rather than removing them. The crowd here was generated for the article; with a real photo, use one you took of people who know about it, or blur the faces that stay recognizable, and treat a stranger's face as something you need permission for rather than something you inherited with the frame. What this article does not cover: background removal, inpainting the subject's edges, upscaling, vertical export, the retro flash-photo look listed alongside this one in the planning brief, and any version longer than the five seconds the clip settings offered.
Written 9 September 2026. Both photographs and all clips were generated in CapCut for this article; no real person, product or brand appears, and the man in them is not a photograph of anybody. Interface labels, model names and cost lines reflect a single session on that date and the configurations named above, and may change. Face size, per-frame pixel change and detail figures were measured on the downloaded files.