You have seen both kinds.
In one, a cat wears a small green cloak and sits in the shadow of a stone hall. A steel pauldron rests on its fur. A brazier lights the edge of one ear and the hem of the cloak. It turns its head toward you, slowly, and the shot looks like a frame from a film.
In the other, the cat is standing. Two back legs, a sword held up in its front paws, shoulders pushing through the cloak, a face caught between cat and person. You cannot say exactly what is wrong, but you know at a glance: that is not a cat.
The difference is not the model version, and it is not luck. It is one decision — whether anyone told the model that the animal must stay an animal.
How it goes wrong
Video models have seen a great deal of human footage.
Say the word "knight" and the strongest association is a person in armour: upright, with shoulders, with hands that can hold a sword. Hand the model a photo of a cat as well, and it now has two conditions to satisfy: the frame must contain a knight, and the subject must come from the photo.
When those conflict, it does not stop and ask. It compromises. It stretches the body, turns shoulder blades into shoulders, reshapes the front paws into something that can grip. That is where the face between cat and human comes from.
There is a second layer. Every template ships with a short reference clip that tells the model how the camera should move. The clip has a silhouette of its own, and the blurrier the animal in your photo, the more the model borrows that silhouette instead.
So a bad render is not a random malfunction. It is a reasonable guess made with too little information. To fix it, you add the missing information.
The first thing we lock
Every template prompt opens with a fixed block. It comes before the scene description, and it is not there to set a mood. It is there to forbid things.
Roughly:
The subject is a pet, and nothing else. Keep the species and breed, the fur colour and markings (including asymmetric patches, stripes and markings around the eyes), the face shape, the eye colour, the ear shape, the body proportions and the tail. Four feet on the ground; never standing on two legs; no human hands, human limbs or human face.
What that does is turn "does it look like my pet" from a matter of taste into a constraint.
The model can still decide how the light falls, how the cloak hangs, how fast the animal moves. It has no room left to turn the animal into a person.
What the lock leaves open
The identity lock only covers what must not change. Everything else belongs to the scene.
The castle knight. Fitted brushed-steel pauldrons, a deep green wool cloak, a weathered bronze clasp, a thin iron circlet. Cold blue moonlight through tall windows; a brazier adding warm light from the other side; faded heraldic banners in the background.
The tower scholar. An indigo velvet robe with worn gold thread at the collar, a soft pointed hat tilted to one side. Candlelight, old books, loose pages hanging in the air, a faint teal light rising from an open grimoire.
The throne. A weathered stone throne with faded velvet cushions, a small tarnished gold crown, a deep crimson mantle with fur trim worn thin. Stained glass glowing behind it, long shadows across the flagstones.
All three share one rule: the costume sits on the animal, it does not grow out of it. And the action is something an animal would do — sitting, turning its head, twitching an ear, a cloak lifting in a draught.
That is also how you judge the result. The costume can be as dramatic as it likes. The body cannot.
Picking a scene
Do not pick the one that looks best. Pick the one that matches the photo.
A pet that sits still with a calm, slightly bored expression belongs on the throne. A pet that tilts its head at everything belongs in the library. A pet with a clear front-on face and readable markings suits the knight, because that shot leans hardest on the face.
If you are unsure, start with the knight. The animal mostly sits still, the light is steady, and it is the easiest result to compare against the photo.
Checking the result
Go from top to bottom. It saves time.
- Points of contact. Four paws on the ground, or a natural sit. If it is standing on two legs, stop there.
- Markings. Compare with the photo: the patch around one eye, the chest, the legs. You are looking for your cat, not a generic tabby.
- How the costume sits. The cloak should rest on the fur with thickness and shadow. If the clothes and the body merge into one surface, the model lost the boundary.
- Hands. There should be none. Paws on a book are fine; fingers are not.
- Motion. Breathing, ears, tail — are the amounts right? Motion that is too smooth, too human, is usually a warning.
When it still misses
If the silhouette is right but the markings are wrong or the face is off, the photo is usually the cause: too dark, too soft, ears cropped, two animals in frame.
Try a clearer photo before you try another style. That is the cheapest order to work in, though it is not a guarantee — sometimes the model or the service is simply having a bad run, and generating again is enough.
If a render fails, or a photo is rejected in review, the credits go back to your balance automatically. You do not have to ask. That part is mechanical, and it has nothing to do with how good the photo was.
The point
What makes a pet knight video work is not "the cat became a knight". It is "the cat is still a cat, and happens to be wearing a knight's clothes".
The first is a visual effect. The second is a joke.
We are making the second one.