Google Flow Consistent Characters: Reference Image Guide

Write a short block covering age, face, hair, clothing, accessories, and traits that must not change. Reuse it verbatim in every related prompt.

Prefer distinctive, visible characteristics over a long list of details that are difficult to reproduce.

Use a sharp image with neutral lighting that clearly shows the face and wardrobe. If the story needs multiple angles, approve a small reference set before video generation.

Do not change the reference, style, lighting, and environment at the same time; controlled changes make drift easier to diagnose.

Generate adjacent scenes with the same core setup and change one variable at a time. Number every shot and store its prompt beside the output.

Test a small group, approve the strongest stills, and use only approved images as image-to-video sources.

Compare facial structure, hair color, wardrobe, body proportions, and props at full size. Small defects in a source image can become more obvious in motion.

If one shot drifts, regenerate it from the closest approved reference instead of changing the entire sequence.

Use a locked order: character name; age range; face shape; skin tone; eyes; hair style and color; wardrobe; accessories; body proportions; and three traits that must never change. Add the action, setting, lighting, and camera for each shot only after this identity block.

Example: Mai, woman around 28, oval face, brown eyes, chin-length black bob, mustard-yellow jacket, dark blue jeans, small silver necklace; preserve the same face, hair, wardrobe, and body proportions in every scene. Scene: walking slowly through a train station, medium tracking shot, soft afternoon light.

When reference features are available, use one clear portrait and one full-body image showing the same wardrobe. Give every reference a clear role; too many images with conflicting lighting or design can cause the model to blend traits.

Create and approve the base character images first, then reuse the same reference set for adjacent shot groups. Change references only when the character intentionally changes wardrobe or story period.

Google Flow is convenient when an approved character image becomes a frame or ingredient for Veo. Kling is worth testing when a project needs several reference types and stronger control. Seedance fits multi-shot progression, although too many changes in one generation can still produce gradual identity drift.

Do not choose a model from one attractive image. Run the same identity prompt and references through three tests: face close-up, full-body shot, and action with camera movement; then score face, hair, wardrobe, proportions, and props.