Flow Ingredients and Scenebuilder: Consistent Characters

A single good clip is easy. A sequence where the same person, wearing the same coat, walks through the same street in five different shots is the real test of a video tool. Google Flow answers it with Ingredients to Video and Scenebuilder. This guide shows how to prepare references, prompt with them, and chain shots into a sequence. Flow is updated frequently, so some labels may have shifted since this was written; the method holds.

Prepare ingredients worth reusing

An ingredient is a reference image Flow uses as a visual anchor. The quality of the result follows the quality of the reference, so spend time here before spending credits.

• Character references: a clear front-facing image, well lit, plain background, the full outfit visible. Add a three-quarter view if you can. Avoid strong shadows, sunglasses and hats unless they are part of the character.
• Object and product references: the item alone on a neutral background, sharp, no reflections of other objects. Two angles help.
• Style references: a frame that shows the palette, grain and lighting you want, without a strong subject that could leak into the shot.
• Keep a small set. Two or three ingredients per shot is manageable; more and the model has to guess which matters.

Whisk, Google’s sibling image tool in Labs, is a quick way to produce clean character sheets and variations. Images from any other tool work too, and the general advice in consistent characters about fixing hair, outfit and distinguishing features applies here as well.

Prompting with Ingredients to Video

Switch the mode to Ingredients to Video, upload your references, and write a prompt that tells Flow how the ingredients come together. The images supply appearance; the prompt supplies everything they cannot show.

1. Refer to each ingredient by a short, fixed description. "The woman in the yellow raincoat" and "the blue enamel kettle", worded identically in every prompt in the project.
2. Describe the action and setting as you would in a text prompt: one action, one camera move, light, and a sound sentence.
3. Do not re-describe what the reference shows. Repeating hair colour and outfit can conflict with the image and cause drift. Mention only what the reference cannot tell the model, such as expression or pose.
4. Keep the style ingredient constant. The same look reference in every shot is what makes a sequence feel like one film.

Generate two outputs for character shots, because likeness varies between them, and pick the one that matches best before you build on it.

Frames to Video for exact starts and ends

When you need a precise opening frame, for example a product in an exact position, Frames to Video is more dependable than ingredients. Upload a start frame, optionally an end frame, and describe only the motion between them. A start frame generated with your character ingredients in Whisk or another image tool gives you both exact composition and a consistent face.

Start and end frames should share a camera position and light. Big differences between them force the model to invent a transition, and that is where faces and objects distort. If you need a big change, use two clips and a cut.

Scenebuilder: extend and jump to

Scenebuilder is where single clips become a sequence. Open a clip you like and you have two main moves.

• Extend continues the shot for a few more seconds from where it ended. The prompt for an extension should describe what happens next, not re-describe the scene. "She turns and walks towards the shop door" is enough. Each extension is a new generation, so expect small drift over several extensions; three or four is a practical limit before you cut.
• Jump to creates a new shot that keeps the scene, characters and style but changes the camera or moment. "Cut to a close-up of her hand on the door handle" or "Jump to a wide shot from across the street". This is how you get coverage of one scene without rebuilding it.

Camera controls in Scenebuilder apply preset moves to a shot or extension, which is handy when you want a consistent push in or pan across several clips without rewording it each time.

Settings at a glance

Keeping a sequence coherent

Even with good tools, a few habits make the difference between a film and a collection of clips. Write a one-line character description and a one-line style description at the top of your project notes, and paste them into every prompt unchanged. Light the whole sequence the same way; a change from dusk to noon between shots is more jarring than a slightly different face. Review the sequence in order after every two or three clips, because drift shows across a cut. Save the ingredient images with the project files, since a prompt without its references cannot be rerun. For the basics of the interface and credits, see Google Flow and Veo 3: getting started.

Try it

Ingredients: [front-facing character reference], [three-quarter character reference], [style frame: cool blue dusk, light rain, 35mm film grain].
Prompt: The woman in the yellow raincoat stands under a bus shelter and checks her watch. Medium shot, camera static, shallow depth of field. Cool blue dusk, light rain, 35mm film grain. Sound: rain on the shelter roof, a distant engine, no music.
Extend: She looks up as headlights approach and steps towards the kerb.
Jump to: Close-up of her face lit by the approaching headlights, rain on her hood.

Run the first prompt, choose the better of two outputs, then apply the extend and the jump-to lines in Scenebuilder. Change the setting sentence for a second scene but leave the character and style wording untouched, and compare how well her face holds across the two locations.