AI Character Generator: Make One Consistent Character

AI VideoCreative

Siamak FreydoonnejadCo-founder & CEO, Sprites.aiPublished on 2026-10-04

The short answer

ChatGPT is the AI character generator most people already have. Describe one character, then ask for a character reference sheet.
For every new image, attach the sheet and describe only the scene. That is what keeps the face, the clothes and the props the same.
For video, Seedance swaps the same sheet into a reference clip. No Higgsfield or Runway in the loop.

Every image here is a real output of ChatGPT's image model, made on October 4, 2026.

Step 1: describe one character

One name, one silhouette, one prop. Nora is a retired pug detective: camel trench coat, tiny felt hat, magnifying glass on a cord, half-lidded skeptical eyes. Specific enough to draw the same way twice, short enough to remember.

Step 2: lock it with a reference sheet

Character reference sheet for Nora, a pug detective in a camel trench coat and felt hat: front, three-quarter and walking side views — *Nora's reference sheet, from one prompt.*

Nora, an original character: a retired pug detective in a camel trench coat and a tiny felt hat, magnifying glass on a cord around her neck, skeptical half-lidded eyes. Make it a character reference sheet: front, three-quarter and walking side profile on one plain off-white background, same lighting and scale in every view. Realistic fur, soft studio light.

Step 3: put the character anywhere

Now attach the sheet as the image input and change only the scene. Four scenes, four prompts, one character:

Nora the pug detective in her trench coat and felt hat under a streetlight on a rainy city street at night — *Rainy street. She stands under a streetlight on a rainy city street at night, puddle reflections, cinematic photo, vertical.*

Nora the pug detective in her trench coat in a red diner booth with a cup of black coffee — *Diner. She sits in a red diner booth with a cup of black coffee, skeptical look, warm morning light, cinematic photo, vertical.*

Nora the pug detective at a cluttered desk examining a photo through her magnifying glass — *Office. She sits at a cluttered wooden desk covered in case files and a desk lamp, examining a photo through the magnifying glass, cinematic photo, vertical.*

Nora the pug detective on a sunny beach holding an ice cream cone, unimpressed — *Beach. She stands on a sunny beach holding an ice cream cone, unimpressed, bright summer light, cinematic photo, vertical.*

Each prompt opens with the same line: "The pug detective from the reference sheet, exactly as drawn: same trench coat, felt hat and magnifying glass." Then the scene. That is the whole technique. Here is the real request for the diner, with the sheet attached:

Terminal screenshot of the OpenAI images edits request that put Nora in a diner, with nora-sheet.png attached as the reference image, and its 200 response — *The diner scene: the sheet goes in as `image[]`. 200 OK in 51.2 seconds.*

What happens without the sheet

Same description, same diner, no sheet attached. Still a pug in a trench coat, but not Nora: the coat turned pale beige, a shirt and waistcoat appeared under it, the hat grew a wider brim and the magnifying glass got bigger and moved to a chain.

Nora in the diner, generated with her reference sheet attached — *With the sheet: Nora.*

A different-looking pug detective in a paler trench coat with a shirt and waistcoat, generated from the same words without the reference sheet — *Without it: a cousin.*

One picture of drift does not sink an ad. Twenty posts of a character whose coat changes colour do: recognition is the reason to have a character at all.

Step 4: make the character move

Video uses the same sheet. Generate a short reference clip in Seedance, prompting the motion and the setting but not the face, then send Seedance one task with the sheet as the reference_image, the clip as the reference_video, and a prompt that says who to replace. Seedance refuses reference videos with real human faces, so make the clip on your own account. The exact JSON and the polling loop are in how to make an AI mascot, and the same recipe is written out per character on every weird AI character page, from Derek Mercer to Tung Tung Tung Sahur.

What it takes

Time: about 35 to 40 seconds per sheet and about 50 seconds per scene on the API, at high quality.
Tokens: Nora's sheet used 6,603 output tokens; the diner scene used 7,349 plus 323 image tokens for the attached sheet.
Tools: ChatGPT for images, Seedance for video. Nothing else.

Want videos on autopilot instead of prompting?

Director by Sprites makes your character's video ads every week: written, generated, reviewed by a human and ready to run. You pick the winners. See Director.

AI character generator questions

What is the best AI character generator?

For a character you will reuse, the one that accepts a reference image. ChatGPT's image model does, in the app and on the API as chatgpt-image-latest: make a reference sheet once, then attach it to every new scene.

How do I make the same AI character in different scenes?

Attach the character's reference sheet as the input image and describe only the new scene. On this page the same pug detective appears on a rainy street, in a diner, at a desk and on a beach that way.

Why does my AI character look different every time?

Because a text description is reinterpreted on every request. The same words gave Nora a paler coat, a shirt and waistcoat and a wider hat when the sheet was left out. The reference image is what holds the details.

Can I make an AI character into a video?

Yes. Generate a short reference clip in Seedance, then have Seedance swap your character in, with the sheet as the reference image. No Higgsfield or Runway needed; the full request is in our AI mascot walkthrough.

How long does it take to generate an AI character?

On the API, about 35 to 40 seconds for a reference sheet and about 50 seconds for a scene built from the sheet, at high quality.