Skip to content
CodeAndBuild LogoCodeAndBuild

AI Video

Keep a Character Consistent Across Google Flow Scenes

Use a reference still, short scenes, and matching prompts so a Google Flow sequence looks like one film.

CodeAndBuild Team8 min read
  • Google Flow
  • Video
  • Continuity
  • Story
On this page
  1. Write a one-page look bible
  2. Generate the reference, then stop redesigning
  3. Extend only while the identity holds
  4. Change one thing per scene
  5. What this workflow is for
  6. Plan sound outside the scene

Google Flow is a scene editor on top of video generation. It will happily give you a beautiful second shot of a different person in a different coat. Continuity is your job. You pick the look once, you feed that still back in as the reference, and you write prompts that repeat the details a viewer uses to recognize someone: coat color, hair, time of day, and the room. You do not ask the model to invent a new lead in every scene.

Treat Flow as a generator of shots, not of a finished 15 minute film. Build the story in scenes of a few seconds, extend only when the last frames are still the same person, and cut in an editor when the face drifts. This guide is for adult characters in ordinary clothes. Do not upload a real person's photo to star as themselves unless you have the right to do that, and do not aim the tool at a minor.

Write a one-page look bible

Before the first generation, write five lines you will paste into every prompt. Who is in the frame, what they wear, where they are, what the light is, and what the camera is not allowed to do. Pin those lines. A look bible that lives in your head will change by scene three because you got bored of the coat.

look-bible.txttext
One adult in a camel wool coat, dark hair tied back, late twenties. Rainy city sidewalk at blue hour, warm shop light on the face, cool street behind. Camera stays at eye level. No extra people, no text, no logos. Match the reference still.

Generate the reference, then stop redesigning

Spend your retries on the still. When the coat, the face, and the street feel like the film, save that frame and mark it as the only reference. Later scenes start from that image or from the last good frame of the previous clip. A fresh text-only prompt for scene two is how you recast the movie by accident.

  1. Scene 1 is the reference. She steps into the shop light and stops.
  2. Scene 2 starts on the last good frame. She turns her head, nothing else moves.
  3. Scene 3 is the same coat from across the street. If the face softens, cut away to the street.
  4. Scene 4 returns to the reference framing so the ending matches the opening.

Extend only while the identity holds

Flow can extend a clip. Use that when the motion is continuous and the face still matches. Stop the moment the jaw, the coat, or the street layout changes. Those frames are not "almost." They are a different shot, and cutting them in teaches the viewer not to trust the person they are watching. It is faster to generate a new four-second shot from the reference than to repair a drifted extend in the middle.

KeepCut awayWhy
Same coat and hairA new jacketThe coat is the costume
One camera moveA second move mid-clipTwo moves read as a jump cut inside the shot
Eye-levelA sudden aerialThe film's grammar changes
Empty sidewalkA crowd that appearsNew faces steal the scene

Change one thing per scene

The prompt delta is the only new sentence. "She looks down at a paper cup" is a delta. Rewriting the weather, the coat, and the lens in the same prompt is a new film. Paste the bible, then add the delta, then repeat "match the reference still." If a scene needs a new location, generate a new reference for that location and accept that you are starting a new look. Do not pretend one still can anchor a kitchen and a rooftop.

Edit in order of the script, not in order of the generations you liked. A gorgeous unused clip is how runtimes explode. Write the voiceover, drop the clips under the lines they illustrate, and export. In the description, say the scenes are generated. Viewers can enjoy a generated film. They should not be told a generated adult is a real documentary subject.

What this workflow is for

  • Pros: a reference still plus a look bible keeps a short sequence recognizable.
  • Cons: long extends still drift, especially around hands and profile views.
  • Pros: cutting back to the master shot hides a weak scene.
  • Cons: Flow will not hold a paragraph of dialogue in sync. Put speech in the voiceover or record it.

Plan sound outside the scene

Flow scenes are pictures. Footsteps, rain, and a line of dialogue will not arrive in sync just because the prompt mentioned them. Record or generate the voiceover as its own track, and cut the picture to it. Ambient rain can be a loop you control. If two scenes should feel like one walk, use the same ambient bed under both so a change in the generated noise does not announce the cut. When a scene has no usable motion, hold the reference still for the length of the sentence instead of extending a clip that has already changed her face. The audience follows the voice. The picture only has to not contradict it.

More guides