A recurring character in your videos
An ambassador who comes back in every video, with the same face, the same haircut and the season's outfit: that's what makes a campaign recognizable. This guide shows how to create that character with Seelab's AI image and video generation, dress it in your products and keep it identical from one shot to the next. Example: Inès, the demo avatar of the fictional brand Maison Ferrande.
1. Create the character, once and for all
It all starts with a stable character. Two methods, from the simplest to the most reliable.
A reference portrait. In Creation, generate a sharp, front-facing portrait on a plain background, with soft light and a neutral expression. That's your "source image". Keep it in a collection and always reuse the same one.
A Seelab avatar model. For a long-running campaign, train a Portraits & Avatars model in Custom models, from a few portraits of the same face. The model learns the features, the haircut and the skin tone: you no longer need to describe the face in your prompts.
On the model page, the Dataset tab (1) shows the training images (2). Use model (3) opens it directly in Creation.

For this demo model, three portraits were enough: all front-facing, on a gray background, with the same light and the same white t-shirt.
Add a character sheet. With the model, generate the same face from several angles. These views then serve as video references when the character turns their head.
Character sheet on a plain light gray background: the same woman in three views side by side, front, three-quarter and profile, waist-up, white t-shirt, soft and even studio light, neutral expression

2. Real people: what the engines accept
Video engines apply their own moderation rules, and they are getting stricter. What to know before launching a campaign:
- Celebrities are blocked. Naming a celebrity or uploading a photo of a well-known person almost always leads to a refusal, whatever the engine.
- Google (Gemini Omni Flash) often refuses real faces. A photo of a real person as a reference can be blocked, even when you hold the rights. A generated face goes through without difficulty.
- ByteDance (SeeDance) no longer accepts photos of real people as references since early 2026. Prefer a generated character.
- Kling accepts the face of a person who has given explicit consent, including for commercial use.
- Minors are excluded everywhere.
And whatever the engine: using the face of a real person (employee, model, influencer) requires their written consent and an image-rights agreement that covers AI. Your legal team remains the right contact.
3. Dress the character in your products
The Multi Refs mode of the Video space brings the character, the clothes and the accessories together in the same shot. The rule: one image per role, and a prompt that says what each image becomes.
- Switch to Multi Refs and add your images from the Resource Hub: the full-length character (Image 1), the trench coat (Image 2) and the bag (Image 3), taken from the datasets of your product models.
- Choose the engine, here Kling v4, in 9:16 and 1080p.
- Write the prompt, naming each image.
The three numbered references (1), the engine (2), the vertical 9:16 format (3) and the prompt that gives each image a role (4):

The woman from Image 1 wears the beige trench coat from Image 2, belted at the waist, and the tan leather bag from Image 3 on her shoulder. She walks toward the camera down a cobbled Paris street, early morning, low golden light. Medium shot, slow tracking shot moving backward with her. Her face, her bob haircut and the details of the trench and bag stay identical to the references.

- A full-length character reference for a complete outfit, a bust reference for a close shot: the framing of the reference guides the engine.
- Sharp product references on a plain background: the dataset photo of your Seelab model is ideal.
- Also describe the outfit in words: color, material, detail ("belted", "gold buttons"). The text confirms what the images show.
- Fewer, better-chosen references: beyond 4 or 5, the engine mixes up the roles more easily.
4. Keep the same character from one video to the next
A campaign is a series of shots. To keep Inès looking like Inès:
- Reuse the same references, in the same order, for the whole series.
- Keep an identical description block at the start of every prompt: "The woman from Image 1, in the same outfit: the beige trench coat from Image 2 and the tan bag from Image 3." Only change what follows: the place, the action, the camera.
- Prefer medium and close shots: a face that is too small in the frame loses fidelity.
- Chain your shots: the last frame of one video can become the start frame of the next.
- Reuse video setup, on the detail view of a successful video, brings back references, engine and settings to launch the next shot.
The woman from Image 1, in the same outfit: the beige trench coat from Image 2 and the tan bag from Image 3. She sits down at a Paris café terrace, puts the bag on the round marble table and smiles as she looks at the street. Waist-up shot, static camera, soft late-afternoon light.

5. Make it speak
With your avatar model, the Custom Models tab of the Video space generates the character with sound. Write the line in quotes, state the language and the tone, and specify what you don't want to hear or see.
The woman, in a black turtleneck, facing the camera in a bright studio with white walls. She smiles and says in French, in a warm, calm voice: "Bonjour, je vous présente notre nouvelle collection." Chest-up close shot, static camera, soft light from the window. No subtitles, no music.

- One short line per shot: 8 to 12 words fit in 4 to 5 seconds.
- Close shot, static camera: lip sync is better when the face is large and steady.
- Same tone, same language from one video to the next, for a consistent voice.
6. Change the outfit without starting over
The shot works, but the collection changes? The Edit tab keeps the motion, the framing and the character, and only replaces the garment. Add the new piece in Multi Refs:
Replace the trench coat worn by the woman with the olive green knit sweater from Image 1, worn with black trousers. Keep her face, her bob haircut, the tan bag, her walk, the setting and the camera movement identical.
The source video (1), the sweater as reference (2), the Kling v4 engine (3) and the instruction (4):


- One change per pass: the outfit, then the setting, then the light.
- Name what stays: face, haircut, walk, framing.
- Check the hands and the face over the whole duration: what you don't mention isn't guaranteed.
Cheat sheet
- A generated character rather than a real person.
- A Seelab avatar model for a stable face across the campaign.
- One image per role in Multi Refs, and each image named in the prompt.
- An identical description block at the start of every prompt.
- Medium or close shots, one action per shot.
- To change the outfit, edit the video rather than remaking it.
With the Seelab team
The Seelab team can create and train your brand's avatar, prepare the models of your products and produce the video campaign for you, agency-style.
Learn more
- Multi Refs: character, product and logo
- Using your model in image and video
- Edit a video
- Video models in detail
Updated on: 12/11/2026
Thank you!
