learnCharacters
How to keep an AI animal character the same in every video
An animal has no face for the engine to lock onto. What goes on an animal's character sheet, how many references each engine takes, and a 6-shot series budget.
by reeldouble · November 2, 2026 · 6 min read
Why animals drift faster than people
Most consistency advice is written for human faces. That is not an accident: video engines are trained on enormous amounts of footage of people, and a face is the thing they have learned to hold steady. Eyes, nose, mouth and jaw give the model a fixed structure to match from one clip to the next.
An animal gives it far less to hold. A ginger cat seen from the side is, to the engine, very close to every other ginger cat. A golden retriever's face in three-quarter view looks like a thousand other retrievers. So the patch over one eye moves to the other eye, the white sock moves from the front left paw to the back right, the dog grows ten centimetres between shots, and the collar changes colour.
The first page of search mostly repeats the human-face method: make one strong reference, reuse it, keep the same description. That is right, and it is not enough for an animal. One guide we read notes that drift grows when the character is small in frame against a busy background. Animals are usually small in frame. The fix is to give the engine more of what makes your animal that animal, and to give it every time.
What goes on an animal's character sheet
For a person, a sheet is mostly the face from a few angles. For an animal, we build the sheet around the four things that drift first.
1. A markings map. Every patch, stripe, spot and sock, and which side it is on. Show the animal from both sides and from above if the back has a pattern. A patch that is only visible from the left must be on a left-side view, or the engine will invent where it goes.
2. Proportions against something fixed. An animal on its own has no scale. Put it next to a plain object in one panel, a chair or a doorway, so its height reads. This one panel stopped our dog changing size between shots more than anything we wrote in the prompt.
3. One prop that never changes. A red collar with a brass tag, a knitted scarf, a saddle. A distinctive prop is easy for the engine to hold and anchors the rest of the animal around it. It also gives the viewer a way to recognise the character at a glance.
4. Flat, even light. Strong shadows on a reference get baked in. A dark band across the back in the reference becomes a stripe in the video. Light the sheet evenly, like a product photo, and let the scene prompt add the mood.
We make the sheet once. In our studio a character sheet costs 30 credits, and after that it is attached to every scene as a reference at no extra cost. If you are starting from a photo of a real pet, our photo to animated character guide covers turning it into a sheet first.
The descriptor block, animal edition
Paste the same short description into every prompt, word for word. For an animal it must name what the sheet shows, in the same order every time:
"Biscuit: small ginger tabby cat, white chest and white front paws, dark patch over the left eye only, short tail with a white tip, red collar with a round brass tag."
Note "left eye only". Words like "only" and "both" stop the engine from adding a second patch to make the face symmetrical. Our descriptor block guide explains why the order of the details matters.
How many references each engine takes
An animal's sheet often needs more images than a person's, because the markings differ side to side. Each engine on our shelf has a hard cap, and the studio enforces it before anything is billed:
| engine | references per render | how you tag it | note for animals |
|---|---|---|---|
| Seedance 2.5 | up to 30 images | @Image1 | room for every side view, plus a second animal and their owner |
| Seedance 2.0 and 2.0 Fast | up to 9 images | @Image1 | enough for a full markings map |
| Kling O3 Pro | up to 5 elements | @Element1 | each element holds a front view plus a reference; use one per character |
| Veo 3.1 Fast | up to 3 images | image 1 | a reference render is always 8 seconds |
For a single animal on its own, Kling O3 Pro holds a character well across shots and has the cheapest silent second we sell. For an animal with complicated markings, or an animal plus two people in the same scene, we move to Seedance, where nine or thirty images leave room for every side view.
Keep the animal big in frame
The single most useful prompt habit: write the shot so the animal fills a good part of the frame. A wide shot of a dog crossing a busy market square is where it turns into a different dog. A medium shot of the same dog trotting past stalls, with the camera following at its height, keeps it.
When a story needs the wide shot, keep that clip short, 4 to 6 seconds, and put a close shot either side of it. The viewer carries the identity through the wide, and if the wide drifts, the retake is cheap.
A 6-shot animal series: what it costs
Animals usually do not speak, so most shots can be silent, and on Kling O3 Pro and Veo 3.1 Fast silent seconds are cheaper. Here is one 36-second episode of a cat story on Kling O3 Pro at 720p:
| line | credits |
|---|---|
| character sheet, made once | 30 |
| six silent shots, 6 s each, at 24 a second | 864 |
| one retake of a 6 s shot | 144 |
| export | 20 (free from Creator up) |
| first episode | 1,058 |
| each episode after | 1,028 |
The same six shots on Veo 3.1 Fast with references cost more than you would expect, because a Veo reference render is always 8 seconds: six of them silent is 960 credits for 48 seconds of footage you will trim to 36. On Seedance 2.0 at 720p, six 6-second shots are 1,620 credits, voiced or silent, but you get nine references per shot.
If the animal has a narrator, keep the clips silent and lay the voice over the top; our narration over silent clips guide covers the setup.
Check every clip against the sheet
Before moving to the next shot, we pause the clip on a frame where the markings are visible and hold it next to the sheet. Patch on the right side? Sock on the right paw? Collar the right colour? If one detail is wrong, retake that clip now, not after the whole episode is built, because every later clip that copies a wrong frame copies the wrong detail too. Our character drift guide lists the fixes in the order we try them.
The quickest way to see whether an engine holds your animal is to test it. Make a sheet in the studio, run one 4-second Kling O3 Pro shot silent for 96 credits, and pause it next to the sheet.
FAQ
How do I keep the same animal in every AI video?
Build a character sheet that shows the markings from both sides, the animal next to an object for scale, and one fixed prop like a collar. Attach it to every scene and paste the same description each time. In ReelDouble a sheet costs 30 credits once.
Why does my AI dog look different in every clip?
Video engines hold human faces better than animals, so markings, size and colour drift. Show every marking on the reference, say which side it is on, and keep the dog large in frame instead of small against a busy background.
Which AI video engine is best for animal characters?
On ReelDouble, Kling O3 Pro suits one animal alone and has the cheapest silent second. Seedance 2.0 takes 9 reference images and Seedance 2.5 takes 30, which suits animals with complex markings or scenes with several characters.
How much does an AI pet video series cost?
On ReelDouble, six silent 6-second Kling O3 Pro shots at 720p are 864 credits. With a 30-credit sheet, one retake and a 20-credit export, the first episode is 1,058 credits and later ones about 1,028.
Can I turn a photo of my real pet into an AI video character?
Yes. Use clear, evenly lit photos from both sides, turn them into a character sheet, and attach that sheet as the reference in every scene so the markings stay put.
