You want a multi-style falling sequence: one man falling while the picture around him turns from a photograph into paper-cut, then pencil sketch, then 3D animation. I built it twice in CapCut from one still, once as separate style clips cut together and once as a chain of clips that each start on the last frame of the clip before, and measured what stayed the same. The man broke before the styles did. In the 3D clip his zipped jacket was open by the third frame, and in the pencil-sketch clip he was a photograph again by the second. The rest of this article is about keeping him one person through the changes, and about where each change goes in a multi-style morph sequence.
The chained version: three 5-second clips, each starting on the last frame of the one before. The GIF plays at 5 frames a second and has no sound.
Make every style from the same still
Everything starts from one still, made with the AI image generator in Image mode (Seedream 4.5, 16:9, 2K). It came back at 2560×1440 with the man upside down in the frame, his face toward the camera and his hands reaching almost to both edges:
A photograph taken from high above, looking straight down at a man in his thirties falling on his back through the open sky, facing up toward the camera. He has short black hair and a short black beard, and he wears a bright red jacket, dark blue jeans and white sneakers. His arms and legs are spread wide. He is in the center of the frame, with empty sky around him. Far below him is a city seen from above, streets and rooftops in soft daylight. No text or logos anywhere.
For the other looks I attached that still in the same composer (the + button, then Upload from device) and asked for a redraw. The paper-cut request:
Redraw this picture as a layered paper-cut artwork made from cut and stacked sheets of colored paper, with visible paper edges and soft shadows between the layers. Keep everything else the same: the man's pose, his position and size in the frame, his face, his short black hair and beard, his red jacket, blue jeans and white sneakers, and the city below him.
The sketch and 3D requests used the same sentence with the style changed: a hand-drawn pencil sketch on white paper in gray graphite lines, and a still from a 3D animated film with smooth rounded shapes. To check that he stayed put, I measured his red jacket in each picture and compared its outline with the jacket in the start still, where 1.00 means the same outline in the same place.
The edits kept his pose, size and place: the jacket measured 841 and 840 pixels wide in the paper-cut and sketch versions against 843 in the start still (1280-pixel-wide frames). The text-only versions didn't. The paper-cut one put the paper around a photographic scene, with a smaller man inside it whose jacket measured 477 pixels wide, and the 3D one raised his arms into a Y and gave him a different face.
The start still and the three edits made from it, with the jacket overlap for each.
Left: edited from the start still. Right: the same styles from text alone.
The sketch took a second try. The first request, in gray graphite only, still left his jacket a faded pink and his jeans blue, too pale for the jacket measure to find. Asking for “only his jacket colored in red pencil” kept a clear red, and that red is the one thing that carries through all four looks unchanged.
What each style clip changed first
Next, each still became a 5-second clip in the AI video generator: Video mode, the still attached, Seedance 2.0 Mini, 16:9, 5s, 720p. The prompt was the same for all four with the look at the end changed:
He keeps falling on his back toward the city below, arms and legs spread, his jacket fluttering in the wind. The camera falls with him and keeps him in the center of the frame. The whole picture keeps looking like a layered paper-cut artwork from the first frame to the last.
All four kept him near the middle of the frame, no more than 49 pixels from where he started, and between 8% smaller and 11% larger. What changed was something else in each one:
The sketch clip is the one to look at closely. At frame 0 it is the drawing. Two frames later the jacket has lost its pencil shading and his face is a photograph, and not quite the same face: the stubble and the line of the jaw differ from the start still. The city stayed drawn. The share of skin-colored pixels in his face went from 0.03 in the drawing to 0.16 by 2 seconds, about the 0.15 of the photographic start still.
The first change in three of the four style clips.
Say what must stay the same on him
I sent the 3D and sketch clips again with one sentence about him added. For the 3D clip: “His red jacket stays zipped up to his chin.” The jacket stayed closed for all 5 seconds, and his jacket outline didn't split into two halves in any frame. The camera moved in closer instead, and the city went out of view.
For the sketch clip: “He stays drawn in pencil like the rest of the picture: his face, hair and hands in gray pencil lines and shading and his jacket in red pencil.” He still changed by the second frame, this time into a colored drawing with inked outlines rather than a photograph. After the first few frames the share of pencil-gray pixels on his torso stayed between 0.28 and 0.43, against 0.12 to 0.18 in the first sketch clip, but his face kept its skin color.
The same two clips at 1 second, before and after the added sentence.
In these two clips, the sentence that named one visible thing, the zipped jacket, held for the whole clip. The sentence that named a drawing technique for his whole figure held only in part.
How long each style can hold before the cut
In a cut sequence, each look runs for a while and then the next clip starts from its first frame. His arms keep moving in every clip, so his pose drifts away from where the next clip begins. I compared his jacket outline at every moment of one clip with the opening frames of the next:
For each moment of one clip, the best overlap with the next clip's first seven frames.
In all three pairs the best match came within the first second. At 0.8 seconds the overlaps were 0.87, 0.57 and 0.79; at 2 seconds, 0.72, 0.40 and 0.60. The paper-cut clip moved his arms the most (its own jacket outline overlapped its first frame by 0.50 after one second), so the paper-cut-to-sketch cut is the weakest of the three for any piece between half a second and 3 seconds long. For the cut version I gave the first three looks about 0.8 seconds each and let the last one run 2 seconds. In the editor:
In the 24 fps export the four pieces came out at 20, 20, 18 and 49 frames, 4.46 seconds in all. The cuts are where it jumps. Measured as the average change from one frame to the next on a 0 to 255 scale, the three cuts came to 50.3, 91.4 and 44.3, against a median of 15.3 inside the pieces, and his jacket outline overlapped by 0.86, 0.51 and 0.81 across them. A hard cut on each change of look can be the effect you want. If it isn't, the next route has no cuts to hide.
The cut version: photo, paper-cut and sketch at 0.75 to 0.83 seconds each, then 2 seconds of 3D (the sketch and 3D pieces are the second-prompt clips). The GIF plays at 8 frames a second and has no sound.
Chain the clips so each joint is one frame
The chained version sends each clip two images: first the frame to start on, second the look to move toward. The first link started on the photo still with the paper-cut still second. Each later link started on the previous link's last frame, which I saved as an image with a frame-grab tool. The request for the second link:
Start on the first image. He keeps falling on his back toward the city below, arms and legs spread, while the whole picture slowly changes into a pencil sketch on white paper, with only his jacket in red, like the second image. The camera falls with him and keeps him in the center of the frame.
Each link opened on its first image, and each look arrived gradually over its 5 seconds rather than at a cut. The looks were approached, not copied. The first link ended as a folded-paper, low-poly city rather than the layered paper-cut of the still. The second turned the streets into a pencil drawing and grayed his face while the jacket stayed red. The third turned the city 3D within its first second and his face into a cartoon by about 3 seconds, then tilted up toward the sky. Checked frame by frame, his jacket stayed zipped through all three links.
I dragged the three links onto the main track in order with no trims (00:15:04 on the editor's clock) and exported at the same settings. The export ran 363 frames, which is the three clips back to back. On the same scale as the cuts above, the two joints measured 16.6 and 14.1, and inside the clips the median step was 8.3 with the largest 20.7. If the next link was generated from the previous clip's final frame, keep that referenced frame at the joint; trimming before it may reintroduce a visible jump.
Average change from each frame to the next in the two exports.
The two edits in the editor: four trimmed style clips, and three untrimmed chain links.
The chained version in nine frames.
What Happened in the Keyframe-Model Test
The composer's model list includes Dreamontage 1.0 Pro Fast, whose settings panel reads “Generates seamless videos with your images as key frames” and offers 3s or 5s of “Duration between frames”. That is a multi-style morph sequence as a single request, so I tried it first: the four stills as images, 3s between frames, 16:9, 720p. In the web session used on September 29, 2026, five Dreamontage 1.0 Pro Fast requests failed and returned the displayed credits. Four used the falling-man stills (full-size PNGs twice, smaller JPEGs, then just two frames), and one used two unrelated stills of a football pitch. Each showed a cost line of 3 credits, followed by “Couldn't generate. Credits returned.”
The composer's reply suggested animating between a first and a last frame with Seedance instead. I tested that with Seedance 2.0 Mini, the photo still and the paper-cut still attached, and “Start on the first image and end on the second image.” The clip started on the first image (a difference of 6.4 from the still on the blurred 0 to 255 measure described at the end) but didn't end on the second: at 5 seconds it was farther from the paper-cut still (68.7) than the two stills are from each other (33.4). It drifted into the folded-paper look instead, and it became the first link of the chain.
Left: the Dreamontage panel (shown with 5s selected; the requests used 3s) and how each request ended. Right: a chained request. The reply below it names Seedance 2.5; the settings line and cost line show 2.0 Mini.
The order to generate in
In this chained workflow, the order is fixed because each new request uses the previous clip's final frame:
For hard cuts instead, generate one clip per look from the edited stills, keep each piece to about its first second, and choose the cut frames where his pose matches.
This article doesn't cover music, cuts on the beat or sound design, the editor's transition effects between looks, or more than four looks. The composer I used showed no style presets, so every look here came from the prompt and the attached image.
Configuration and cost. Stills: Image mode, Seedream 4.5, 16:9, 2K, where the composer's cost line read 1 credit for each; seven in all (the start still, four edits including the gray sketch, and two text-only versions). Clips: Video mode, Seedance 2.0 Mini, 16:9, 5s, 720p, with one or two images attached, where the cost line read 25 credits for each clip; nine in all (four style clips, two second-prompt clips and three chain links). The account's credit details for September 29, 2026 list seven charges of 1 credit and nine of 25, 232 credits in all, and the balance in the header went from 24,421 to 24,189. Six requests failed: five Dreamontage 1.0 Pro Fast requests (16:9, 3s between frames, 720p, cost line 3 credits) and one Seedance 2.0 Mini request that succeeded on a second send. The credit details show a line of +3 for each Dreamontage request and +25 for the other, with no matching charge lines, and the balance change above matches the sixteen completed generations. Nothing done in the editor, including the two exports, 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 and the clips 1280×720 at 24 frames per second, 121 frames. Each generated still and clip carries a small “Ai” mark in its top-left corner. The two exports are 1920×1080 at 24 frames per second, 107 and 363 frames; in both, the video starts 0.083 seconds after the audio. The cut version's GIF is 560 pixels wide at 8 frames a second; the 15-second chained version's GIF is 448 pixels wide at 5 frames a second to stay under 4 MB. Prompt text is quoted as sent. The composer's replies to at least four of the Seedance 2.0 Mini requests named Seedance 2.5; the settings line under each request and the cost line read 2.0 Mini and 25 credits.
How the jacket was measured. Strongly red pixels (the red end of the hue circle, saturation above 90 on a 0 to 255 scale), keeping every piece at least 8% the size of the largest so that an open jacket still counts as one jacket, and leaving out pieces that touch the frame edge. Overlap is the shared area divided by the combined area. A planted red bar read back its planted area and position, and a copy shifted 10 pixels gave the expected 0.714.
How the changes were timed. Face: the share of skin-colored pixels in a box just below the jacket, where his head is. Pencil shading: the share of low-saturation pixels in a box on his torso. Jacket opening: the frame where the red outline first splits into two large pieces. City: the share of orange-red roof pixels away from the man. The frame numbers in the table were then checked on full-size frames.
How the cuts and joints were measured. The mean absolute difference between consecutive exported frames over all three color channels, at 640×360, on 0 to 255. Each exported frame was matched to its source clip frame by the smallest difference. The start-and-end check on the first link used the same difference on frames blurred slightly to ignore fine texture. The cut curves compare the jacket outline at each frame of one clip with the best of the next clip's first seven frames.
Corrected along the way. My first decode of the exports padded the 0.083-second gap at the start with two copies of the first frame, which made the exports look as if they repeated their opening frame; decoding with the file's own timestamps shows no repeated frames, and the frame counts above come from that. The first jacket measure kept only the largest red piece and lost half the jacket once it opened, and in the text-only paper-cut still it picked up red paper at the frame edge; counting all large pieces and dropping pieces that touch the edge fixed both.