The first clip is the easy one. You write a pair of animals, you get a pair of animals, and it looks like a series is about to happen. The trouble starts on episode two, when the same sentences come back with a different animal in them and nothing ties the two clips together.
This piece builds three episodes of an AI character video series at 16:9 in CapCut, moving the same felt rabbit and hedgehog through three rooms. Three ways of carrying the pair from one episode to the next were run against each other, and the numbers below come off the exported clips.
Three episodes cut together: 1280 by 720, 30 frames per second, 15.17 seconds, with a track from the editor's audio library. Shown here as frames in sequence; the full video and a GIF version are supplied alongside this document. The rabbit, the hedgehog and all three rooms were generated for this article.
The second episode is where an AI character video series breaks
A single clip has no continuity to fail. The moment there is a second one, the audience has something to compare against, and small differences in fur tone, face shape or size read as a different animal rather than the same animal in a new room.
So the question this article answers is narrow: given a first clip you are happy with, what do you hand the generator for the second one? There are three obvious answers, and they do not behave the same way.
Lock the character in text before the first clip
Before anything is generated, the pair gets written down once, as a block that stays fixed from then on. Six things are worth fixing, because each of them is something a later generation can quietly alter: size relative to each other, fur color and material, facial proportions, a fixed accessory each, posture, and emotional baseline.
Two small felt animal characters. The rabbit is dove grey, made of soft felt, and sits upright on its hind legs; it is about twice as tall as its friend, and it wears a red knitted scarf wound twice around its neck. The hedgehog is warm tan with a cream belly, also sitting upright, and it wears round brass glasses. Both have large round black eyes about a quarter as wide as their faces, and small dark noses. Both look calm and friendly. Handmade felt texture, visible fibers, matte surfaces, no shine.
Two accessories are doing real work there. A red scarf and brass glasses are cheap to write, easy to see, and easy to check later: either the scarf is in the frame or it is not. That makes them the first thing to look at when an episode comes back.
Each episode then adds two paragraphs to that block and nothing else — one for the room, one for the event. Episode one put the pair on a rug in a living room in the evening, and had the hedgehog look up while the rabbit leaned down.
The settings line and the cost line for an episode request. The mode chip has to be set to Video by hand; left on Auto it picked a different model, which is covered further down.
That request came back at 1280 by 720, 24 frames per second, 121 frames. The clip is usable, and more importantly its first frame is a clean front-on view of both characters, which is what the next episode needs.
What the reference frame actually locks
The obvious move for episode two is to save a frame from episode one and attach it to the next request. In the CapCut composer an attached image becomes the first frame of the generated clip, and that is a stronger constraint than it sounds like.
Episode two was written as the same character block plus a kitchen and a cup being passed across. With the reference frame attached, the pair came back intact — and still in the living room.
The same episode-two request, sent three ways. Only the bottom-right route reached the kitchen with the pair intact.
The room can be measured rather than argued about. Taking a block of wall that holds no character in any of these clips, and comparing it with the same block in episode one, gives a single number per clip.
How far each clip's wall block sits from episode one's. The reference-frame route reads 2.9, which is the same distance episode one's own frames sit from each other.
At 2.9 the reference-frame clip is no further from episode one than episode one is from itself. The kitchen in the prompt was ignored. That is the cost of attaching a frame: it carries the character and the set together, and in this run the set won.
Describing the pair again is not the same as showing them
The opposite move is to send the character block as text with no attachment, which leaves the generator free to build the kitchen. It does. The room reads 8.6, comfortably a different room.
What arrives with it is a different rabbit. The declared fields that can be read straight off the pixels tell the story.
Averages over three frames of each clip. The description-only column is the one that moves.
The scarf survived every route, which is the argument for writing a hard accessory into the block. The fur did not. Episode one's gray reads 0.266 in brightness; the description-only kitchen reads 0.477, close to twice as light. The comparison is fair because the two kitchen clips were asked for in the same words, so the difference is the route rather than the room.
Three things changed that a measurement does not capture, and they are worth naming as things seen rather than counted: the rabbit's ears and muzzle are drawn differently, the hedgehog's spine pattern is finer and more even, and its glasses are larger. The scarf, the glasses and the upright posture are all still there. It is recognizably the same idea and not the same pair.
Move the room in a still, then animate it
The two failures point at the fix between them. The reference frame holds the characters, so keep using it — but spend one cheap step moving them into the new room as a still image before asking for any motion.
That step is an image request with the reference frame attached, and it is worth writing narrowly: name what to keep, and say that the room is the only thing being replaced.
Keep the two characters exactly as they are: the same dove grey felt rabbit with the same red knitted scarf wound twice around its neck, and the same warm tan hedgehog with the cream belly and round brass glasses. Keep their sizes relative to each other, their upright poses, their positions in the frame and their faces unchanged. Replace only the room around them: they are now on a wooden kitchen worktop in the morning, with a plain tiled wall behind them and daylight coming from a window on the left. No other changes.
The room-change step runs in Image mode with the reference frame attached. The cost line reads one credit at these settings.
The still that came back. The pair is carried over; the room is new.
Attaching that still to the episode-two request gives a clip that reaches the kitchen and keeps the animals. The room reads 25.6 and the fur brightness reads 0.299, against episode one's 0.266 — a third of the drift the description-only route produced.
One thing does change that no one asked for. Moving the room also moved the camera back: the pair sits smaller in the frame than in episode one. If framing matters to you, it is the thing to check on the still before spending a video credit on it.
Episode three, and whether the drift piles up
Two episodes prove very little. The real question for a series is whether each new episode drifts a bit further, so that episode ten has nothing to do with episode one.
Episode three went to a study with a shelf of books, using the same two-step route — and starting from the original reference frame rather than from episode two. Anchoring every episode to the same saved frame is what keeps the chain from compounding.
Episode three's gray reads 0.377 in brightness, further from episode one than episode two was at 0.299. So the drift is not zero, and a long series will want a fresh look at the character block every few episodes. It is still well inside the description-only route's 0.477, and the scarf, the glasses, the relative size and the poses all came through.
Both clips and the still were made through the AI video generator and AI image generator flows, then cut together in the web editor. Three clips dropped in together land on the main track in filename order, so naming them a1, a2, a3 puts the episodes in sequence without any dragging.
The three episodes on the timeline with a library track underneath, trimmed to the same length.
What the eight requests cost
One of those requests deserves its own note. The first attempt at episode one went out with the mode chip left on Auto, which selected a different and much more expensive model. The cost line read 185 credits, the generation ran to 99 percent, and then it failed.
The failed request returned its credits. Setting the mode chip to Video by hand put the composer back on the cheaper model at 25 credits.
The cost lines as shown under each request, read on 14 September 2026.
A hundred and twenty-seven credits were charged in the end, of which fifty went on the two comparison clips that exist only to make this article's argument. Building the same three episodes without them costs 77. No cost line appeared for the uploads, the library track, the trims or the export. These figures are for these configurations on this account and this date and come from the line the composer shows before each request; other models, durations, sizes and modes are charged differently, and prices change.
The file the series actually depends on
Everything above rests on one saved PNG. The character block is retyped each time, the rooms and events change, but the thing that makes episode seven look like episode one is a single frame pulled out of the first clip.
The saved frame every episode starts from. Both characters face the camera, neither is covered, and the lamp behind them is warm but soft.
Three things make a frame worth saving, and all three come from what the later steps do with it. It has to be front-on, because the room-change step redraws whatever is hidden. It has to have both characters unobstructed, because anything covered gets invented. And it should come from a moment of even light rather than a strong colored one, because that color travels into every room you move them to afterward. The frame used here was the very first frame of episode one, before any motion had started.
Save it next to the character block in the same folder, and treat the pair of them as the series bible. When a new episode looks wrong, the first thing to try is going back to that frame rather than to the clip that just came back.
What is not covered here: voice-over and captions, style presets, image enhancement, 4K export, transitions between episodes, and any real animated character or brand, none of which appear in this piece. The rabbit, the hedgehog and the three rooms are invented for this article, and the music is a royalty-free track from the editor's own library.
Written 14 September 2026. Every clip and still here was generated in CapCut for this article and assembled in its web editor; no real character, brand or person appears. Interface labels and cost lines reflect a single session on that date and the configurations named above, and may change. Room distance is the mean absolute pixel difference between a fixed upper-right block of each frame and the same block in episode one; scarf and fur readings are color measurements taken inside a window anchored to the scarf, averaged over three frames of each clip. Changes in ear shape, spine pattern and glasses size are described from looking at the frames, not measured.