How to keep the same face in every AI shot: the recipe that worked

By Deep Bhalerao, founder · Published

Deep, Nora and Otto talk and smile together at night in the rain.

The first AI clip anyone makes usually looks great. The second one has a slightly different person in it. By the tenth, your hero has a new nose, a new shirt and a new age. That drift is the single biggest thing that stops an AI film from feeling like a film.

In OVERCLOCK, my kaiju series made in Agent Social, the same people appear across four episodes and dozens of clips. I play the lead myself. Here is the recipe that kept our faces the same, and the mistakes that taught it.

Rule one: make each character once

The rule in my cast document is one sentence: make each one once, then attach the same picture to every take they appear in. Not a new portrait per scene. Not a description that the video model has to imagine again every time. One picture, reused everywhere.

For myself, I started from my own photos. For the rest of the cast the studio painted them: a face close-up and a turnaround, which is the full body seen from the front, both sides and the back, on a plain grey background. Each character also gets a written voice line: age, tone and accent, so they sound the same too.

Reference portrait of Deep: a young man with dark curly hair and a white T-shirt.
The reference portrait of Deep that every OVERCLOCK Episode 1 shot of him reused.

Rule two: one identity sheet per character, per look

Agent Social glues each character's pictures into a single wide identity sheet. The face close-up sits in a fixed box on the left. The rest of the sheet shows the full body from each side: front, left, right and back for a person, and front, back and side for a creature. Nothing is painted again to make it. The sheet is stitched in code from pictures that already exist, so the face on the sheet is exactly the face you approved.

Why one sheet instead of several pictures? Every video model has a limit on how many reference pictures one clip can take. Before the sheet, each person used two of those slots. In a scene with two people, that was four slots gone before the room or the previous shot got any. With the sheet, each person needs one.

  • Each outfit gets its own sheet. If a character changes clothes, that is a new look with its own sheet, and the face stays the same.
  • To change a character, change the source picture. Ask for glasses and the studio edits the close-up, then rebuilds the sheet. You never edit the sheet itself.
  • Takes that are already filmed keep the pictures they were made with. Changing a character never quietly rewrites an old clip.

Rule three: attach it to every take, by name

A take is the written text for one clip. Every take lists the pictures it uses, and the text refers to each person by name with a tag, so the video model knows which picture is which person. If a character is in the take, their sheet is attached.

One-off extras are different. A waiter who appears once is described in that take's words and gets no sheet. Anyone who appears in two or more clips gets a picture.

Places need the same care as faces

A face that holds is wasted if the room behind it changes. In OVERCLOCK every place starts as one photo. The main places were then filmed as a full turn on the spot, so every take set there can pick the view that faces the way its camera faces. The café looks like the same café from the door and from the counter. Vehicles are places too: Deep's SUV is a place picture, never a prop.

Deep, Nora and Otto at a hotel breakfast buffet, Deep holding a plate.
Episode 3, a hotel breakfast. Same face as the reference portrait, in a different place, light and weather.

Moments: a still of the shot before you film it

The last part of the recipe is the moment. A moment is a still picture of one beat of the story, painted from the character sheets, the place pictures and the last frame of the previous clip. It is not the first frame of the clip. It is a reference that shows the video model what this beat should look like: who is where, facing which way, in which room.

Moments are also where I learned the most. One painted moment was meant to look deeper into a shop. It came back with the whole room turned about ninety degrees so the glass front stayed in view, and a white van copied in from a photo that faced the other way. The causes were all in how the request was put together, so the studio was changed:

  • The staging, meaning who is where and what is happening, now leads the request.
  • When a moment continues from an earlier picture, it has to say whether the camera keeps its view or moves.
  • Every place picture sent with it says what it is there for.
  • A moment is checked against the pictures it was made from before it can be used, and only checked moments are reused later.
  • Left and right are spelled out for people who face the camera, after one moment came back with them swapped.

The whole recipe on one card

  • Make each character once: face close-up plus turnaround, or your own photos.
  • Check the pictures before filming with them.
  • Let the studio glue them into one identity sheet per look.
  • Attach that sheet to every take the character is in, and tag them by name in the text.
  • Give every recurring place a picture, and every camera direction its view.
  • Paint a moment for the important beats, check it, then film.

None of this needs a camera or an actor. It needs patience with pictures before the first render, which is much cheaper than patience with footage after it.

Related: Consistent characters in AI video — same faces every shot

Make your own film in Agent Social

Sign up
How to keep the same face in every AI shot: the recipe that worked — Agent Social