Consistent character AI: the same face in every image and video

Why faces change between generations

Image and video models start from noise and a prompt. Without a reference, even a very detailed description gives a different person on each run. The hair is close, the face is not, and a feed made of those pictures looks like five different people.

Reference photos give the model the real face to follow. That is the whole idea behind a saved character: the face comes from photos, and your prompt only describes the scene.

Meet Tara

Tara is one character, made from a short description and three portraits. Everything below is a different scene, outfit and light, made with the same saved character and nothing else changed between runs.

Look at the face, not the background: the shape of the eyes, the brows, the nose, the jaw and the way the hair falls. Those hold from the cafe to the rain to the city at night.

What a character is made of

A character in Sparebar is a saved set of things that are sent to the model together with your scene.

• Reference photos. Up to three clear photos of the same person, or a portrait generated from a description. This is what carries the face.
• A look profile. Gender, age range, skin tone, hair length, colour and style, features such as earrings or a nose pin, and everyday wardrobe. These fill the gaps a photo cannot, and they are repeated in every prompt.
• Extra angles. One button adds a three quarter and a full body shot of the same person. Close ups use the front shot, outfit shots use the full body one.
• Saved outfits. A name, a description and an optional photo, so the same jacket or saree can appear again later.
• Niche and tone of voice. These shape the captions Sparebar writes for the character.

Same character, four different products

Consistency matters most in ads, where a viewer decides in a second whether a person looks real. Here is Tara in four ads for four unrelated products: a serum, a water bottle, a coffee pouch and a lip tint. The product, scene, outfit and headline change every time. Her face does not.

Read more about how these are made on the product ads page.

A second character, built the same way

Nothing here is special to Tara. Riya has wavy brown hair and a different face, and she holds together across a balcony, a cafe and a bedroom in the same way. You can keep as many characters as you like and switch between them in one click.

Different people, different looks

Meera and Kabir. A man and a woman, different ages and skin tones, from the same flow: describe, pick a portrait, save.

Consistency in video

Video models that accept an image start from your character photo, so the person in the clip is the person in your library. For talking videos Sparebar makes the still first, in the same scene, and then animates it. That keeps the face and the setting steady from the first frame.

• Talking head tip: she speaks a line you type.
• Product pitch: she holds a saved product and talks about it.
• Day in my life: a short, moving moment from a routine.

What you can change without losing the face

This is the useful part. Almost everything about a shot is yours to change, and the character stays put.

• Scene and place: a cafe, a street, a bathroom, a balcony in the rain.
• Outfit: type it, or use a saved outfit.
• Product: pick a saved product with its photo and label.
• Photo or video, shape (9:16, 4:5, 1:1, 16:9), resolution and length.
• The model: a fast one for tests and a stronger one for the final.
• The words: the headline on an ad, or the line she says out loud.

Tips for better consistency

Faces stay very close, not pixel identical. Models differ, and a few simple habits help a lot.

• Use sharp, well lit photos that show the face from the front, then add the angles.
• Keep the scene description about the scene. Let the photos carry the face.
• Fill in the look profile. It repeats the details a photo cannot show.
• Use a stronger model for hero shots. Seedance 2.0, Veo 3.1 and Nano Banana Pro follow references closely.
• Close ups and busy scenes drift the most. If a run drifts, run it again, it is cheap, and your credits are refunded when a generation fails.
• Some video models refuse photos of real people. If one does, try another model.