The advice around anthropomorphic animal clips is that they fail messily — that the animal comes out deformed, so you should generate a batch and keep the one that holds together. Two clips at 25 credits each say the failure is quieter than that and needs a different response.
Both came back clean. Good light, steady framing, a convincing ginger cat wearing headphones, no melted limbs, nothing to throw away on quality grounds. Both also show a cat walking on four legs, which is not what the prompt asked for.
A ginger cat walks upright on its two hind legs along a plain bedroom floor toward the camera, wearing small over-ear headphones, front paws swinging loosely at its sides. Its hind paws stay flat on the floor and carry the weight on every step. Plain bedroom wall behind, nothing else in the frame, no people. Locked-off tripod shot at cat height, no zoom, no push-in, no dolly, no pan and no crop. No text, no logos, no printed words and no brand marks anywhere in the frame. 16:9 landscape.
The instruction is stated twice, once as the action and once as a weight-bearing constraint. Neither clip followed it. This is the part worth knowing before spending: a batch-and-select workflow catches deformity, and deformity is not what went wrong. Every clip in the batch would have passed a quality check and failed the brief.
The same sentence, asked of a different mode
The same posture was then requested from the AI image generator as a still, three ways of wording it, 1 credit each. All three stood up.
Scored against a rule fixed before generating: body vertical, weight on the hind paws, front paws off the floor.
The wording was not the problem and neither was the model's ability to picture the pose. Asked as a still, the same request lands on each run. Asked as a clip, it is quietly replaced with the thing cats usually do. That is a property of which generator is being addressed, not of how carefully the sentence is phrased.
Buy the pose in stills, then animate it
If the still holds the pose and the clip does not, the pose should be bought where it is cheap and carried into the clip as a reference rather than re-requested. The best of the three stills went through image to video on Seedance 2.0 Mini at 16:9, 5 seconds and 720p, 25 credits, asking only for movement and explicitly not for the posture.
The cat stays standing upright on its two hind legs the whole time and bobs its head and shoulders to a beat, front paws swinging a little at its sides. Its hind paws stay planted on the floor. 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.
1280 by 720, 24 frames per second, 121 frames, 5.04 seconds.
The cat stays upright for the whole clip. Checking every fourth frame across the 121, the posture holds in all of them: body vertical, hind paws planted, front paws swinging clear of the floor. In this test, supplying the upright pose in the reference image produced the desired posture, while the two text-to-video attempts did not.
Which gives the working order, and it inverts the usual advice about batching:
Batching still matters. It just belongs on the 1-credit step rather than the 25-credit one. Three stills cost 3 credits and settled a question that two clips at 50 credits had not.
What to write in each of the two prompts
The division of labor between the two prompts turns out to be clean, and worth stating plainly because it is easy to get backwards.
The still prompt owns the body. Posture, proportions, what the animal is wearing, where its weight sits. The phrasings that worked described the skeleton rather than the activity: spine vertical, shoulders above hips, hips above the hind paws, front legs hanging at the sides like arms. One that also worked gave the cat a reason to be upright — reaching for something above it — which produces a stretched, balancing stance rather than a standing one, so pick according to which you want.
The video prompt owns time. What moves, how far, and what stays put. It should restate the posture as something to maintain rather than to achieve, and it should name the parts that must not move. The clip above says the hind paws stay planted, which is the difference between a cat bobbing on the spot and a cat that drifts back onto four legs halfway through.
What it cost
103 credits against the cost lines, of which 75 went on the route that did not work. The article you are reading is the argument for spending 3 instead. One of the three initial clips could not be retrieved from the workspace after generating, so it is counted as spent and is not shown; the two that downloaded both behaved the same way, and a third clip agreeing with them would not have changed the recommendation.
Ask the image model for the body and the video model for the clock
The reflex when a clip comes back wrong is to rewrite the clip prompt. Two attempts here say that particular instruction was not going to survive in that particular place, however it was phrased — and that the same words work immediately one step earlier, for a fortieth of the price.
So the question to ask of a failed clip is not "how do I say this better" but "is this something the video model should have been asked for at all". Posture, costume, proportion and identity belong to the still. Motion, duration and what holds still belong to the clip. When something refuses to appear in a clip, try asking for it as a picture first — it costs one credit to find out, and if the picture has it, the clip will keep it.
Written 16 September 2026. Every frame and clip was generated in CapCut for this article; no footage or photograph of a real animal was used. Clips are 1280 by 720, 24 frames per second, 121 frames, 5.04 seconds; stills are 2560 by 1440 at 16:9 and 2K. Posture was scored by eye against a rule fixed before any generating: the body axis vertical, the weight on the hind paws, the front paws clear of the floor, held for the majority of the clip. The frames behind every verdict are published above at the resolution they were judged at, so the scoring can be checked rather than taken on trust. Three automatic measures were built and discarded before that. A step-cadence figure from the autocorrelation of the silhouette's floor-band width returned correlation strengths of 0.19 and 0.12, too weak to carry the implied tempo, which makes sense once the cat is seen walking toward the camera rather than across it. A deformity proxy from frame-to-frame silhouette area jumps was tracking floor reflections and cast shadows rather than limbs. A standing-versus-walking measure from the bounding box's height-to-width ratio returned figures that contradicted the pictures, because on these backgrounds — a bed, a floor edge, a shadow — the largest non-background component is not reliably the animal. The common cause is that automatic subject extraction does not hold up on furnished rooms, which is why the categorical question was answered by inspection with the evidence published instead.