How to keep products and characters consistent in AI video

A viewer will forgive an unusual camera move. They notice when the bottle changes shape or the presenter becomes another person between cuts.

Consistency begins before video generation. The motion model cannot protect details that were never fixed in the reference set.

Decide what must remain identical

For a product, list proportions, cap or closure, label layout, logo, color, material, and any legally required text. For a recurring person, fix face, approximate age, build, hair, clothing, and distinctive details.

Keep the list short enough to review on every shot. “Same product” is too vague.

Build a reference sheet

Use clean views from the angles the film needs. Include front, side, three-quarter, close detail, and scale. A character sheet should include neutral and useful expressions under consistent lighting. A product sheet needs packaging details at a resolution where a reviewer can compare them.

Do not ask the video model to invent the unseen side of an important object unless invention is acceptable.

Approve the still before adding motion

The input image defines the starting composition, subject, lighting, and style in an image-to-video workflow. Runway's current image-to-video guide, which is optimized for Gen-4.5, therefore tells creators to use the text prompt mainly for motion.

That division is useful. Fix appearance in the reference image. Ask the video prompt for action, environmental movement, camera behavior, and timing.

Change one variable at a time

Start with the essential movement. Add camera motion or environmental effects after the subject behaves correctly. If the face changes after a dramatic orbit is added, the cause is easier to identify.

Complex prompts hide the source of failure. They also make a lucky result difficult to reproduce.

Protect difficult details

Small packaging text and logos remain fragile. A production can avoid some failures by limiting extreme angles, keeping the product large enough in frame, shortening the shot, or adding exact typography during editing rather than asking the model to regenerate it.

That is not cheating. It is compositing.

Review at cuts, not only frames

Compare the last clear frame of one shot with the first clear frame of the next. Check face shape, hairline, clothing seams, product height, label position, light direction, and the side of the frame where the subject stands.

A shot may be internally convincing and still fail beside its neighbor.

Keep a rejection vocabulary

Use repeatable reasons: identity drift, proportion change, label change, logo corruption, wardrobe change, lighting discontinuity, camera error, hand error, or physical impossibility. The log will show which shots and models consume the budget.

“Looks off” is hard to fix.

When to stop regenerating

Use ordinary footage when the product must be exact through a complex movement. Use a still with controlled camera animation when the label matters more than object motion. Composite the product into a generated environment when that provides better control.

The goal is a reliable cut, not proof that every pixel came from one model.

Sources and limits

• Runway Gen-4 video prompting guide
• Runway image-to-video prompting guide
• Runway camera terms and examples

The exact controls differ by model and may change. No workflow eliminates drift in every shot.

Sources were checked on October 21, 2026.