Measuring this forced a word apart. "The same character" turns out to be two different claims: the same character as a viewer recognizes it, and the same character as a pixel-level repeat. A locked description gets you one of them reliably and the other only while nothing else in the frame moves. Which one you were counting on decides how you write the description.
The character, written as three things
The character is one anomalous material, one fixed accessory, and a face simple enough to describe in six words. Everything else about it is free to change.
The locked description. A short round character made entirely of corrugated cardboard, knee height to a person. The fluted edges show where the sheets were cut, and the brown paper surface is scuffed and slightly dented. It has two small black dot eyes and no mouth. A small brass bell hangs from a red cord around its neck.
Against it, the line somebody would actually type: em>a funny cardboard mascot character standing in a kitchen at night/em>. Both went to the AI image generator three times each, at 16:9, with no other changes between runs.
Three runs each. The top row has the bell in all three. The bottom row has it in none, and the character is a different design each time — ears in one, a cap in another, a different head shape in the third.
What the measurement says
Each set was measured against its own first run: a patch of the character — eyes, cord, bell, torso, no background in it — searched for across the other two frames at every scale between 0.6 and 1.45.
1.000 would mean the character came back pixel for pixel. The loose set scores lowest because the search cannot find the character in the other frames at all — its best guess lands on a kitchen cabinet.
Locked, with the scene held identical: 0.508. Loose: 0.131, and the box lands on background rather than on anything belonging to a character. That is the difference between a design that comes back and a design that gets reinvented each time you press the button.
Then the scene changes, which is what a series is
Episodes are not all set in the same kitchen. So the same locked description went out twice more with the setting swapped — a laundromat, then a bus shelter — and nothing else touched.
The character arrives. It is cardboard, it has the two dot eyes, the bell is on its red cord. Measured, it scores 0.187: the search finds it, and the match is a third of what it was when the kitchen stayed put.
Both of those are true at once, and that is the whole finding. The traits repeat. The drawing does not. The fluting runs at a different angle, the head is rounder in one and squarer in another, the bell hangs at a different length. A viewer scrolling past reads one character across all five pictures. A template match reads five different renderings of a shared idea.
One episode, moving
The first run then went to the AI video generator as the opening frame of a five-second episode.
Five seconds from the first still. The material, the eyes and the bell hold through the movement.
Within one clip there is nothing to worry about: the episode is generated from a picture of the character, so the character in it is that picture. The consistency problem lives between episodes, not inside them, which is a useful thing to know about where to spend effort.
Writing the three elements so they survive
If the drawing is going to be redone each episode, the three elements have to be things that survive being redrawn. That rules out more than it sounds like.
The catchphrase is the one element that costs nothing to keep. It is not in the image at all, so nothing can redraw it — it goes in the caption, or through text to speech on the same voice each episode, and it carries the series identity for free while the picture does the harder half.
Where this stops
Eight runs, one character, one model. Three runs per condition is enough to separate 0.508 from 0.131 and not enough to put a tolerance on either. The new-scene figure rests on two runs.
The measure is narrow in a way worth saying plainly: it rewards a pixel-level repeat, which is not what a viewer is doing. It is useful here because the three conditions are ordered the same way a viewer would order them, not because 0.187 means anything on its own.
Pick traits that can be redrawn
A design that only works when it is reproduced exactly is the wrong design for this format, because exact reproduction is the thing that does not happen between episodes. The material, the one colored accessory and the six-word face all came back in every run that asked for them, in three different rooms, drawn differently each time. Choose the character out of things that can be said in a sentence and recognized from across a room, and the redrawing stops being a problem to solve and becomes the thing you were relying on.
Configuration and cost. All eight stills: Image, Seedream 4.5, 16:9, 2K, Standard — the panel declared 1 credit per send. One send did not register and consumed nothing; it was repeated. The clip: Video, Seedance 2.0 Mini, 16:9, 5s, 720p, Standard — the panel declared 25 credits. Total 33 credits, read from the send panel on each send, 2026-09-21. Credit costs are tied to that configuration on that date; other models, resolutions or durations are priced differently.
Files.t1.png, t2.png, t3.png (locked description, same scene), l1.png, l2.png, l3.png (loose line), s2.png and s3.png (locked description, laundromat and bus shelter), all 2560×1440; ep1.mp4 1280×720, 24 fps, 121 frames, 5.04 s, shown on this page as ep1.gif.
The prompt block is reproduced as it was sent.
How the numbers were taken, and what was thrown away. The match is a multi-scale normalized cross-correlation of a character patch, run after local contrast normalization of both patch and frame. It was self-tested twice: against known rescalings of one frame, recovering 1.00, 0.86 and 1.26 for true 1.00, 0.85 and 1.25; and against known lighting changes — gamma 1.8, gamma 0.6, brightness at 0.45 and contrast halved — where it returned 0.995 to 1.000. That second test also ruled out the explanation I first reached for: the plain version of the same matcher is equally lighting-invariant, so the drop across scenes is not a lighting artifact. The plain version was discarded for a different reason: with the scene changed it lost the character altogether, putting its best match on a washing machine and scoring the same character lower than a different one. Two color measures for the accessory were also built and discarded — both counted 42 to 58 percent of the character region as red, because warm cardboard under warm light is red in RGB — so the bell is counted from the published frames instead, where it is visible or not. Every match box was drawn back onto its frame and checked before any figure here was taken from it.