AdverForge✦

REFERENCE IMAGE · PRODUCT VIDEO

Reference Image to AI Product Video: A Practical Ecommerce Guide

Learn how to choose product reference images, define allowed motion and review AI product video outputs without losing product identity or inventing unsupported claims.

By AdverForge10 min read
The short answer

Choose a reference that clearly defines the product, plan only motions supported by that evidence, keep text out of the generated footage and review the result against the same source image before using it as an ad.

REFERENCE SELECTION

The source image is the product contract, not decoration.

A stylish image is not automatically a useful product reference. The model needs enough visible evidence to preserve the item while the camera and scene change. Start with the cleanest image that defines the exact product version, then add verified secondary views only when the planned shot needs them.

Identity

Is identity clear?

The full product silhouette, finish and control layout are visible.

Coverage

Can the planned angle be supported?

The image shows every surface needed by the planned camera angle.

Separation

Is separation clear?

The product is not hidden by hands, props, packaging or heavy reflections.

Evidence

Is evidence clear?

Any feature shown in motion is supported by the product page or supplied assets.

If the product is small, obscured or shown only inside a lifestyle scene, obtain a better reference before generation. Prompt length cannot recover visual evidence that is not present.

MOTION AND SHOT CONTROL

Describe what may move and what must stay locked.

ByteDance documents reference-based generation as a way to control subjects, motion, camera work, lighting and other visual properties. For a product ad, that flexibility still needs a narrow brief: one advertising job, one approved action and one continuity rule.

  1. 01

    Choose one advertising job

    Use the shot for recognition, feature proof, a verified use moment or desire. Do not mix every purpose into one clip.

  2. 02

    Name one approved action

    Specify a camera push, orbit, controlled product movement or supported interaction. State its starting and ending frame.

  3. 03

    Lock product identity

    Preserve the silhouette, finish, controls, materials and approved accessories visible in the primary reference.

  4. 04

    Separate footage from copy

    Generate the product shot first. Add prices, claims, captions and calls to action after the footage passes review.

Use the product consistency prompt templates to turn those decisions into reviewable fields.

REAL ADVERFORGE CASE

One clean coffee-machine reference becomes one controlled product moment.

Unbranded coffee machine reference image used for an AI product video
Input referenceThe body, controls, spout and finish define the product identity.
Reviewed outputA simple use moment keeps the machine as the hero and avoids generated prices, logos or claims.

This is not evidence that every generation will remain consistent. It shows the review standard: the output should still be recognizable as the submitted item, and the scene should not claim more than the reference and approved plan support.

OUTPUT QA

Compare the result with the evidence, not with the prompt alone.

Identity
Does the silhouette, finish, control layout and accessory set still match?
State
Did the product perform only the approved action or transition?
Claims
Did the shot add unsupported performance, compatibility, certification or copy?
Readability
Can a viewer understand the intended product or feature without narration?
Assembly
Is there enough stable framing and safe area for captions and the final CTA?

Review more finished clips and their shot logic in the AI product video examples, or follow the complete Seedance ecommerce workflow.

REFERENCE IMAGE FAQ

Common questions before image-to-video generation.

What makes a good reference image for an AI product video?

Use a sharp image in which the product is large, unobstructed and visually complete. The silhouette, materials, controls and approved accessories should be easy to inspect.

Can one reference image keep a product perfectly consistent?

No. A strong reference reduces ambiguity, but the output still needs review. Use additional verified views when a shot exposes details that the primary image does not show.

Should text and logos be generated inside the video?

Usually not. Generated text can drift or become unreadable. Keep critical copy, prices and calls to action in the editing layer after the product footage passes review.

How many product actions should one short clip contain?

Prefer one advertising job and one approved action per short shot. More actions create more opportunities for the model to change the product or imply an unsupported feature.

For current reference capabilities and limits, see the official Seedance 2.5 launch notes.