This article builds a 35-second beach and vacation reel of one AI-generated cat from seven clips in CapCut, and measures what held and what broke along the way. Seven clips allow 42 possible cuts, each the last frame of one clip against the first frame of another, and I measured the color difference at all 42. Thirty came to between 5 and 23. The twelve that touched the one clip set on a pier at sunset came to between 33 and 47, and none fell between 23 and 33.
The cat was the easy part of these cat AI videos. Every clip started from a still made in CapCut, and with the same base still attached to each request, the cat's head carried over from scene to scene in this run. What changed from scene to scene was the cat's size in the frame, the camera, the first frames of each clip and the color of the light. The sections below take those four in that order, then the title and the order of the clips.
The finished reel: seven five-second clips of one AI-generated cat, each from its own still, with a warm shift on the six daytime clips, the pier clip last and a white title with a black stroke. The GIF is a sample, frames 36 to 95 of each clip at every fourth frame, played at 6 frames a second with no sound, so its cuts are not the cuts in the video file.
One base still, one changed sentence for each beach scene
The base still came from the AI image generator in Image mode. The settings line read Seedream 4.5, 16:9 and 2K, and the composer's cost line read 1 credit. This is the prompt:
A photograph of a white-sand beach at midday: turquoise sea behind, a few palm trees at the edge of the frame. In the center of the frame stands a fluffy gray tabby cat upright on its hind legs, facing the camera, with both front paws raised. It wears round black sunglasses and a bright red bandana around its neck. The whole cat is visible from ears to feet. Camera at the cat's eye level, medium wide shot. No text anywhere.
The cat is described by three marks that are easy to look for again later: gray tabby fur, round black sunglasses and a red bandana. Every other scene is an edit of this still. In Image mode, click + in the composer, choose Upload from device and pick the still; it appears as a chip at the start of the request. The request repeats the three marks in its first sentence and changes only the scene sentence, shown here in brackets:
Keep the same cat, the same round black sunglasses and the same red bandana. The cat now [scene sentence]. The whole cat is visible. Camera at the cat's eye level, medium wide shot. No text anywhere.
The scene sentences below swap one of four things: a prop, an outfit, a place or an activity, and the jet ski, the banana boat and the lounger each bring their own prop with the activity. The last column is a count of head features that match the base still's head, explained further down.
For the boat and the jet ski, the visibility sentence named the vehicle as well: “The whole cat and the whole banana boat are visible.” The shirt and hat request had to drop the bandana from its first sentence, so it read: “Keep the same cat and the same round black sunglasses. Replace the red bandana with a yellow Hawaiian shirt and a straw hat. The cat stands on the same white-sand beach with both front paws raised. The whole cat is visible. Camera at the cat's eye level, medium wide shot. No text anywhere.”
To see what the base still adds, I ran three of the scenes again with no still attached, describing the cat in the base prompt's words. This is the jet ski version; the banana boat and lounger versions differ only in the vehicle sentence:
A photograph of a white-sand beach at midday: turquoise sea behind, a few palm trees at the edge of the frame. A fluffy gray tabby cat with round black sunglasses and a bright red bandana around its neck sits on a white jet ski with blue stripes on open turquoise water, holding the handlebars with both front paws. The whole cat and the whole jet ski are visible. Camera at the cat's eye level, medium wide shot. No text anywhere.
The base still and eight more. Each label says how the still was made, and the number is the count of head features that matched the base still. The lounger's head is tipped back, so it has no number here.
To ask whether a still shows the same cat, I cropped each head around its two lenses, scaled the crops so the lenses match in size, and counted the local image features that agree with the base still's head. I tested the count on planted cases first: the base head enlarged by 15%, shifted, with its contrast and noise changed, matched with 603 features, and the head mirrored left to right matched with 13.
With the base still attached, five of the six scenes matched with 33 to 264 features (table). With no still, the three scenes matched with 3 to 19: the banana boat with 6, the jet ski with 19 and the lounger with 3. The lounger edit matched with 7, in the range of the text-only stills, but its head is tipped back and seen from below, which the count does not follow, so I compared that one by eye. In the base still, both ears carry a red rim that no prompt asked for. The pier, jet ski, banana boat, coconut and lounger stills carry it too; the hat covers the ears in the sixth, and the three text-only stills have no red on the ears. The two text-only stills in the figure below also show a fluffier cat and thinner sunglasses frames than the base still's. In this run, attaching the base still is what kept the head, and a description alone did not.
The head in the base still, in two edits and in two text-only stills, with the shirt and hat edit at bottom right, where the hat covers the ears.
Props changed the cat's size in the frame; the outfit did not
The cat's size in the frame is not in the sentence, so I read it from the sunglasses: the diameter of a lens in a 1,280-pixel-wide copy of each still, divided by the base still's 58.2 pixels. A planted test with dark circles 52 and 64 pixels across read 52.0 and 64.0.
The outfit and the place left the size where it was: 1.00 for the shirt and hat and 1.03 for the pier. Giving the cat something to hold, ride or lie on moved it: 1.85 for the coconut, 1.16 for the jet ski, 0.87 for the banana boat and 0.81 for the lounger. Every request asked for a medium wide shot and for the whole cat to be visible, and in the coconut still the frame ended at the cat's belly. With no still attached, the banana boat, jet ski and lounger scenes read 0.63, 1.15 and 1.29, against 0.87, 1.16 and 0.81 with it, so attaching the base still did not fix the size in this run.
When I wanted the paddle already in the banana boat still, I attached the boat still instead of the base still and asked for the paddle. That edit read 0.88 against the boat still's 0.87. If the size matters for a cut, look at the still before you spend a clip on it; at the settings listed at the end, a still is far cheaper than a clip.
The cat's size in the frame, read from the width of the sunglasses, for the edits (blue) and the text-only stills (orange).
Paddling pulled the camera back by about a tenth
Each clip started in Video mode of the AI video generator with its own still attached: click + in the composer, choose Upload from device and pick the still. The settings line read Seedance 2.0 Mini, 16:9, 5s and 720p, and the composer's cost line read 25 credits for the request. Each request was one sentence for the action and one for the camera:
The first take of four of the clips at 0, 1.7, 3.3 and 5 seconds.
Every sentence ends with the camera staying still. To see whether it did, I followed the sea, sky and palms frame by frame. Corner features outside a tall box under the cat's head were matched between neighboring frames, a scale and a shift were fitted to them, and the fits were chained back to the first frame. A background scale of 0.90 means the scenery is drawn 10% smaller than in the first frame. On a planted test, a push-in to 1.100 with a drift of 40 and −25 pixels and a 3-degree roll read back as 1.100, 40.7 and −25.2 pixels and 2.98 degrees. A second method, which compares the distances between pairs of features at frame 12 and at the end, read a planted 0.900 as 0.900.
The two takes of the dance sentence ended at 0.998 and 0.996, and the two takes of the lounger sentence at 0.999 and 0.997. The pier clip ended at 1.001. The shirt and hat clip drifted in to 1.015, and the coconut clip stepped back to 0.983 in the first second and ended at 0.980.
The paddling did not hold. In take 1 the background scale read 0.992 at 1 second, 0.972 at 2, 0.946 at 3, 0.922 at 4 and 0.897 at 5; take 2 ended at 0.889. An 8-second request with the same sentence ended at 0.907 after 8 seconds and read 0.963 at 4. With the paddle already in the still and the words “with a small paddle” dropped from the sentence, the clip ended at 0.878. With the cat sitting still while the boat floated, it ended at 0.985. The four clips in which the cat paddled lost 9 to 12 percent of the scenery's scale, and every other clip I could follow changed by less than 2.5 percent. The second method agreed with the first within 0.012 on eleven of the twelve clips; on the paddle-in-the-still clip it read 0.921 against 0.893.
Background scale over each clip, first frame = 1.00, for the twelve clips whose background could be followed. The two jet ski clips are left out.
The paddling clip (take 1) with its time and its background scale, measured against the first frame; the overlay was added for this article. The GIF plays at 8 frames a second and has no sound.
The two jet ski clips are not in the chart. With only sea and sky in the frame, the second method kept 8 and 7 features, and its reading for the first clip, 0.986, sat 0.03 from the first method's 1.016, so I report no number. In the frames I looked at, the ski and the cat sit in the same place while the spray behind them grows.
Water and vehicles are where I expected the pictures to break, so I looked for the cat's paws or the paddle passing through the boat or the ski. In nine frames from the first paddling clip (1.25 to 5 seconds) and nine from the first jet ski clip (0 to 5 seconds), the paws stayed on the paddle, the hull and the handlebar, and I did not see anything pass through. Eighteen frames are not every frame. The first two seconds of the paddling clip are the closest to its neighbors' framing, with the scale still at 0.972 after 2 seconds. I did not try to correct the pull-back in the editor.
Twelve of thirteen clips started three frames lighter than the rest
The first frame of each daytime clip is close to its still. I measured lightness as the L* of the mean color of the frame below its top fifth; frame 0 read between 1.0 and 1.4 below the still's. Three frames later it had fallen to between 7.8 and 12.5 below. In the twelve daytime clips the fall from frame 0 to frame 3 was 6.8 to 11.3, by frame 3 each clip was within 1.9 of the level it held afterward, and frames 12 to 59 averaged 8.3 to 11.3 below the still. The pier clip, whose still is dark (L* 43.2), moved by 0.6 between frame 0 and frame 3 and settled 0.9 below its still.
At a cut, the first frame of the next daytime clip sat 7 to 10 L* above the level it settled at, and it fell to that level within three frames. In this run, each cut into a daytime clip carries 0.125 seconds of a lighter picture. Cutting the first three frames off a clip takes that away, as the trimmed clip below shows. I did not trim the seven clips in the reel. I computed the six cuts again to each next clip's fourth frame, and the sums fell by 13 to 25 percent (next section). To see that trimming works in the editor, I trimmed one clip:
The exported copy had 118 frames instead of 121, and its first frame read L* 63.0, against 71.9 for the untrimmed copy's first frame and 63.9 for the level both settled at.
Mean L* of each clip's frame below its top fifth, minus the same measure on its still, for the first 25 frames of thirteen clips.
Matching the stills widened the cuts, and a warm shift narrowed one
The clips came back darker and less colorful than their stills. Their settled level was 8.3 to 11.3 below the stills' in L* for the six daytime clips and 0.9 below for the pier, and their chroma, the distance from gray in the a* and b* channels, was 0.25 to 0.79 of the stills'. The number I use for a cut is the straight-line distance between the mean colors of two frames in CIE L*a*b*, taken over the part of the frame below its top fifth (ΔE76). To learn what the editor's numbers do, I exported six copies of the dance clip in one reel (one of them with its first three frames trimmed, used in the last section) and read the level each settled at, the mean from frame 12 to the end:
In this calibration, Brightness raised L* and lowered chroma, and Saturation raised chroma and left L* nearly where it was. The dance still read L* 73.7 and chroma 5.6, against 63.9 and 3.0 for its clip. I chose Brightness 14, a little under the calibration's 20, and Saturation 35, between its 20 and 40. I set them on the first clip and clicked Apply to all at the bottom of the panel, which put them on the other six clips as well, the pier clip included. The six daytime clips then came within 2.0 L* of their stills, and their distance from their own stills fell from between 8.5 and 11.9 to between 2.4 and 5.3.
The cuts got worse. The sum of the six cuts in the order I built went from 84.9 with no grade to 109.9. The five cuts among the daytime clips rose from 52.1 to 65.2, and the cut from the banana boat to the pier rose from 32.9 to 44.7. The stills, cut against each other in the same order, differ by 45.2 over the five daytime cuts and by 44.6 at the banana boat to pier cut. So the grade left the pier cut where the stills already had it, and put the daytime cuts above the stills' own. In this run, matching each clip to its own still did not match the clips to each other.
Temperature moved the other way. With Temperature set to 30 or 50 on the six daytime clips and the pier clip left at 0, the sums came to 80.6 and 73.4. The five daytime cuts barely moved (51.0 and 50.2, against 52.1). The change was in the cut from the banana boat to the pier, 29.6 and 23.2 against 32.9; at +50 the boat clip's settled chroma had risen from 3.5 to 19.9, toward the pier clip's 28.4. The Temperature box would not take more than 50: I typed 60 and it read 50.
Two cautions. At +50 the six daytime clips sat 21.7 to 25.4 from their own stills, with the chroma of five of them raised to between 13.9 and 19.9 (the jet ski's stayed at 9.8), so the cut narrowed because the pictures got warmer than the stills; look at the reel above before you copy the setting. And the no-grade numbers come from the clip files, while the graded ones come from exports of the editor. An ungraded segment of the dance clip in the calibration export read within 0.2 L* of its file, but the pier clip, which was at 0 in every reel, read 2.8 from its file, mostly along the blue-to-yellow axis. That is about as large as the 3.3 the +30 setting took off the boat-to-pier cut, so I count +30 as no better than no grade and +50, which took 9.7 off, as an improvement.
Left: the color difference at all 42 possible cuts, from the clips as generated. Right: the sum of the six cuts in the order I built, for four ways of grading, measured to each next clip's first frame (solid) and to its fourth frame (hatched, computed from the same exports).
A black stroke held the title at 5.5 or better on all seven clips
The reel needed one title that reads over seven different pictures. In the editor, Text and then Add heading put a heading on the timeline. I typed COCONUT COVE, set its size to 8 (the size box offered 5 to 32 and started at 16) and, under Transform, moved it to X −500, Y 430, near the top left of the frame and clear of the small “Ai” mark that each clip carries in that corner. Then I dragged the right edge of the title's bar to the end of the last clip, so one title covers all seven.
To find out whether white letters can be read there, I exported the reel with and without the title, took the difference to find the letters' pixels, and computed the contrast between the letters and the picture under them as (L + 0.05) / (Lb + 0.05) on relative luminance. Sliding the same letter footprint to other places in the untitled export gave four placements to compare. At the top left, where I put it, the contrast was 5.1 to 6.1 on the six daytime clips and 2.2 on the pier clip. The bottom left gave 1.4 to 3.1 on the six daytime clips and 7.9 on the pier clip, the center gave 2.8 to 6.9, and the top right gave 3.2 to 5.4 on the daytime clips and 1.9 on the pier clip. The worst clip at each placement read 2.2 at the top left, 1.4 at the bottom left, 2.8 at the center and 1.9 at the top right.
A black stroke fixed the top left. In the text panel's Basic section I clicked the swatch on the Stroke row and chose black. The contrast between the letters and the pixels right around them, which is what the stroke controls, was 1.8 to 3.2 without it, lowest on the pier clip, and 5.5 to 5.7 with it on all seven clips.
Where the title goes, and what the stroke does at the top left.
The title without and with the black stroke, on the pier clip and on the dance clip.
The cat stayed clear of the title in the frames I checked. In each clip I took the frame in which the cat's head came closest to it, and in all seven the title and its stroke are clear of the head.
The frame of each clip in which the cat's head came closest to the title, with the title and its stroke boxed.
Put the pier clip last
The grid on the left of the figure in the color section shows why the pier clip is the one to place. Every cut into it or out of it read 33 to 47, and every other cut read 5 to 23, with a median of 11.6. The banana boat clip is the next most different: cuts leaving it read 16 to 23, against 5 to 14 for cuts leaving the other daytime clips.
I searched all 5,040 orders of the seven clips for the smallest sum of the six cuts. In the clips as generated, the 100 cheapest orders all ended with the pier clip and none began with it. Orders with the pier last (720 of them) summed to 75.6 through 111.4, with the pier first (720) to 87.8 through 109.1, and with it in the middle (3,600) to 106.9 through 144.4. The first order I tried, with the pier second, summed to 133.9. The order I built, with the pier last and the other six in the order the GIF shows, summed to 84.9, and 68 of the 5,040 orders were lower. The best order (jet ski, lounger, dance, coconut, shirt and hat, banana boat, pier) summed to 75.6, and the best order in each of the four grading conditions, with and without the first three frames of each clip, also ended with the pier clip.
A clip at either end of the reel is touched by one cut and every other clip by two. The smallest step into the pier clip came from the banana boat, 32.9, and the smallest step out of it went to the banana boat, 37.8, so at the end the pier clip adds one cut of 32.9 and at the start one of 37.8 or more. Where the pier goes moved the total by 31 (75.6 against 106.9 for the best order with it in the middle); the order of the other six moved it by about 9 (75.6 against the 84.9 I built).
The reel in the editor:
The requests behind a clip and an edit, and the editor panels used for the grade, the title and the export.
Before you place your own clips, lay each clip's last frame beside every other clip's first frame. If one clip stands apart from the rest, as the pier clip did here, put it last, where only one cut touches it.
This article doesn't cover vertical 9:16 output, clips longer than 8 seconds, more than one cat in a scene or the clips' own sound. Its numbers come from one cat and seven scenes made on September 30, 2026, with second takes of the dance, lounger, banana boat and jet ski scenes to check the first. I did not measure the camera in the jet ski clips, did not build a reel with the first three frames trimmed and did not check the cat's paws against the title frame by frame.
Configuration and cost. Stills: Image mode, Seedream 4.5, 16:9, 2K, where the composer's cost line read 1 credit for each. Eleven were made: the base still, six scene edits, one edit of the boat still to add the paddle and three text-only stills; all eleven appear in the numbers above. Clips: Video mode, Seedance 2.0 Mini, 16:9, 720p, where the cost line read 25 credits for each of thirteen 5-second clips and 40 credits for the one 8-second clip. Seven of the fourteen clips are in the reel, the first take of each scene; the others are second takes and variants used for the camera and color numbers. The account's credit details for September 30, 2026 list eleven charges of 1 credit, thirteen of 25 and one of 40, 376 credits in all, and the balance in the header went from 23,575 to 23,199. Nothing done in the editor, the exports included, 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. Each still and clip carries a small “Ai” mark in its top-left corner. The clips are 1280×720 at 24 frames per second with their own sound: 121 frames (5.04 seconds) for the thirteen 5-second clips and 193 for the 8-second one. The export is 1920×1080 at 24 frames per second, 847 frames, and its picture starts 0.083 seconds after its sound. The reel GIF is 480 pixels wide at 6 frames a second, the setting for long reels, and the boat GIF is 560 pixels wide at 8 frames a second. Prompt text is quoted as sent. COCONUT COVE is an invented place name.
How the head and the size were read. The two lenses are the two near-black, near-round blobs side by side in a frame, and their mean diameter is the ruler. On a planted 120-frame sequence with the lenses hidden in frames 30 to 39, a bandana hidden in frames 60 to 69 and look-alikes around (a dark bar, a lone dark blob, red socks, two small dark paw pads), the tracker flagged the same frames as planted and placed the head with no error. The head match crops 3.6 lens diameters to each side, 3.2 above and 4.0 below the lenses, scales the crop so a lens is 60 pixels, evens out its contrast and matches SIFT features (ratio test 0.75); the count is the number of matches that agree with one scale, rotation and shift within 6 pixels.
How the camera was read. For each pair of neighboring frames, up to 600 corner features outside a tall box under the cat's head were followed, a scale, rotation and shift were fitted with outliers rejected at 1.5 pixels, and the fits were multiplied back to frame 0. The second method follows features in the upper half of the frame outside a 600-pixel column around the cat, keeps those that track forward and back, and takes the median ratio of the distances between pairs of features more than 150 pixels apart. Its planted test used 99 features and 3,668 pairs.
How color was read. Every color number is the mean CIE L*a*b* of the frame below its top fifth, from 640×360 copies of the frames, and differences are ΔE76. Planted checks: two frames whose lower parts differed by 10.0 read 9.89 even with the top fifth painted differently in one of them, and a difference of 5.0 read 5.05; a planted lightness drop (L* 70, 66, 62, then 60) read 9.7 between frame 0 and frame 3 and a settled level of 60.2. The settled level is the mean over frames 12 to 59 (frame 12 to the end in the calibration table), and chroma is the square root of a*² + b*².
How the title was read. The letters' pixels are where the export with the title differs from the export without it by more than 60 in some channel, plus pixels within a few pixels of those that differ by more than 12. On a planted white box the method found its bounding box to the pixel and read 1.52:1 on a pale background and 5.39:1 on a dark green one, each equal to the value worked out independently, and the same footprint moved 700 pixels read 5.39:1 again. The contrast around the letters, with the stroke, is the letters' luminance against the ring of pixels within 3 pixels of them; a planted white bar inside a 10-pixel black frame read 21.0:1 against 21.0:1 worked out independently. Six frames per clip, 16 frames apart, were averaged.
How the orders were searched. Each cut is the last frame of one clip against the first (or fourth) frame of another, from the built reel's exports for the graded conditions and from the clip files for no grade. All 5,040 orders were scored by the sum of their six cuts. On a planted matrix in which one order's six cuts scored 1 each and every other cut scored 10, the search returned the planted order. The first-three-frames sums are computed from the same exports and were not built.