Write "one of them is icy and one of them is fiery" into a video prompt and the generator has two ways to read you. It can take icy and fiery as colors, and hand back a white animal and an orange one. Or it can take them as materials, and build one body out of ice and the other out of flame. Those are very different clips, and which one you get is decided by how much you wrote.
This piece writes a pair of ai animal characters as a four-row sheet, folds the sheet into a prompt block that stays fixed, and then sends the same five-second request four times, changing one thing each time. Building the pair in the first place, and running it as a series, is covered separately in Fire and Ice Penguin Couple; what follows is the test of whether the sheet is doing anything.
The pair the sheet produced: 1280 by 720, 24 frames per second, 5.04 seconds. Shown here as frames in sequence; the full video and a GIF version are supplied alongside this document. Both characters were generated for this article.
Four rows are enough to define an ai animal character
A single animal in a clip has nothing to play against. A pair has a relationship, and a relationship is a format you can keep using rather than a one-off. What a pair needs before you write anything is four decisions, taken once per character.
The sheet behind the four clips in this piece. Build is deliberately identical for the two characters so that the other three rows carry the contrast.
Build is the silhouette, and keeping it the same for both characters is what makes the two read as a pair rather than as two unrelated creatures. Material is the row that does the most work, and the one this piece spends the most time on. Marking is where the material shows from a distance. Pace is the only row that is about behavior, and it is the row the shot has to make room for.
Two blocks, and only one of them moves
Written out, the sheet becomes one paragraph for the AI video generator to read. The shot becomes a second one. Keeping them as separate blocks is what makes the rest of this testable: change the cast and hold the shot, or change the shot and hold the cast.
Block 1, the cast. Two 3D cartoon penguin characters standing on a snowy ledge, rounded bodies, big glossy eyes, soft studio lighting. FROST stands on the left: his body is pale frosted translucent ice, blue-white, with a dusting of snow across his shoulders and a short crest of clear icicles on his head; he is calm and slow. EMBER stands on the right: his body is warm orange glowing ember, with small flames along the crest of his head and the tip of his tail; he is quick and eager.
Block 2, the shot. They face each other, lean in slowly until their noses touch, hold for a moment, then lean back to where they started. Their feet stay in the same place and nothing else in the scene moves. Plain pale blue sky behind them, no other characters, no text and no logos. Cartoon 3D animation. Locked-off tripod shot, no zoom, no push in, no dolly, no pan and no crop, and the last frame has the same framing as the first frame.
Four requests, each changing one block. The four clips used one configuration throughout: Video mode, Seedance 2.0 Mini, 16:9, 5s, 720p.
What "icy" means when you do not say more
The second request replaced Block 1 with a single sentence — the same two penguins on the same ledge, described only as "one of them is icy and one of them is fiery" — and left Block 2 untouched. What came back is two ordinary fluffy chicks, a white one with a small snowflake on the chest and a rust-colored one.
Ember pixels is the share of a body whose CIELAB b* reading is above 45 — the range the glowing orange sits in. Readings are averaged over all 121 frames.
The warm character went from 77.9 percent ember-saturated pixels down to 4.8 percent. Its median b* fell from +58.7 to +33.7 and its median lightness from 162 to 113. That is the difference between a body that emits and a body with orange feathers on it. The cool character moved much less, from a median b* of −17.0 to −14.5, because a pale blue-white animal and a body made of ice land in a similar place on that axis.
The crest is where the material shows
The other half of the answer is in the outline. Counting prominent peaks along the top edge of each body, the sheet version's ice character carries 3.2 of them per frame — the crest of icicles — and its ember partner 2.3, which are the flames. In the version without the sheet, the same counts are 0.0 and 0.6: two round heads.
This is the practical reason the Marking row earns its place. At the size a clip gets watched, the spiky crest is what tells a scrolling viewer the animal is made of something, and it is the first thing that disappears when the prompt only names a temperature.
The action came out the same either way
Measured on the upper bodies only, so the snow they stand on does not bridge the two.
Both clips performed the lean-in. The pair from the sheet closed a 263-pixel gap and made contact at 1.29 seconds, holding it for 51 of 121 frames. The pair without the sheet closed a 287-pixel gap, made contact at 2.29 seconds, and held it for 36 frames. The timing differs, the action does not.
That is a useful separation to have proved rather than assumed. Block 2 is carrying the action; Block 1 is carrying what the actors are made of. When a clip disappoints you, the two blocks are the first place to look, and they fail in different ways.
The same sheet on a different animal
The anchor for this format is penguins, but the sheet does not care. The third request changed the Build row from penguin to fox and left every other word of Block 1 and all of Block 2 alone.
Frame 1 of each clip. Only the Build row and the species word changed between them.
The materials transferred. The ember fox reads 53.2 percent ember-saturated against the ember penguin's 77.9, and carries the same 2.3 crest spikes; the ice fox keeps its icicle crest at 2.2 spikes and the snow across its shoulders. The lean-in happened too, contact at 1.54 seconds and held for 52 frames. Foxes are the only substitution tested here, so treat the Build row as the row to vary and the rest of the sheet as the part that travels.
The same pair in a second scene
The fourth request kept Block 1 word for word and replaced Block 2: the two sit side by side facing the camera, EMBER waves a flipper, FROST looks over and back. Nothing in the shot asks them to touch, and they do not — the gap holds between 95 and 118 pixels for the whole clip, which is the flat line in the chart above.
The same instant, 1.25 seconds in, from each of the four clips.
The materials came along. The ember body reads 63.7 percent ember-saturated with a flame still standing on the crest, the ice body keeps its icicle crown at 2.2 spikes and the snow across its head, and the cool character is on the left in all 121 frames. What moved is everything the sheet does not mention: both characters are drawn rounder, with larger eyes, and the readings all slide toward the middle — the ice body from a median b* of −17.0 to −9.0, the ember body from +58.7 to +50.0.
What the four clips cost
The configuration chips sit directly above the price they belong to.
One hundred credits in total, of which fifty went on the two clips that exist to make the comparison. The balance in the header read 24,191 before the first request and 24,091 after the fourth, which matches the four cost lines. A pair in one scene is 25. All four went through the AI video generator flow from text alone, with no image attached and nothing to upload. These figures are for this configuration on this account and this date, read from the line the composer shows before each request; other models, durations and resolutions are charged differently, and prices change.
The row your sheet does not have yet
Four rows held up across a species change and a scene change. But two things changed that the sheet says nothing about. The icy fox came out 19 percent taller than its ember partner while the two penguins sat within 6 percent of each other, and the second scene redrew both characters rounder and wider-eyed. Relative size and proportion are not on the sheet, so the generator decided them, differently each time.
That is the method for growing the sheet, and it costs nothing. Put two of your own clips side by side and list what differs that you did not write. Each item on that list is a row you are missing, and adding it is one more line of Block 1 rather than a new request. A pair that holds its material across four clips is already a format; the rows you add after that are what keep it the same pair from one post to the next.
Both characters here were written from scratch for this article and generated from text, so nothing was uploaded: no photograph and no real animal went into them. If you start from a design that belongs to someone else, that is a different question and not one this piece answers. The composer's own line under the send button asks you to use only media you have rights and consent for, which is worth reading before you attach anything.
The penguin is a carrier, not the format. Foxes are the one substitution tested here and they worked; otters, rabbits and the rest are untested, and the way to find out is to change the Build row and leave the other three alone.
Also not covered: audio, clips longer than five seconds, resolutions above 720p, style presets, camera-movement settings, and cutting several of these together into one post.
Written 14 September 2026. All four clips were generated in CapCut for this article and are 1280 by 720, 24 frames per second, 121 frames, 5.04 seconds. Interface labels and cost lines reflect a single session on that date and the configuration named above, and may change. Each body was isolated by comparing the frame against a background estimated from its own left and right edges, then split at the column where the two upper bodies are furthest apart. Ember pixels is the share of the eroded body mask whose CIELAB b* is above 45; median b* and lightness are taken over the same mask; crest spikes counts peaks on the top outline with a prominence of at least 12 pixels after a light blur; the gap is the widest run of background columns between the two upper bodies. All readings are averaged over 121 frames unless stated otherwise.