The phrase ‘use reference media’ hides several distinct operations. An image can set the opening frame, a collection of portraits can guide a character’s appearance, a source clip can supply motion or camera language and an ending image can specify the final composition. Choosing the wrong reference type can waste an entire batch of videos while giving a false impression of model inconsistency.
Reference types in plain language
An opening image is a composition anchor. It is most useful when the first frame must match a product photograph or drawn scene. An ending image constrains where the shot should arrive. A character reference tries to preserve identity across shots without locking every camera angle. A video reference may communicate gesture, body movement or shot rhythm. No one of these automatically guarantees the others.
Google’s Veo 3.1 documentation historically described up to three reference images, frame interpolation and video extension. Its three Gemini API preview identifiers were later shut down in October 2026. Luma Ray3.14’s official guide says it does not support character reference even though it offers keyframe and modification workflows. ByteDance Seedance 2.5’s July announcement discusses large image/video/audio reference bundles. The lesson is to read the exact model reference table and deployment surface, not a generic social-media demonstration.
Match the input to the problem
For a costume that must remain fixed, first create approved look references and use an engine with a documented identity mechanism. For a dance performance that needs the same timing as licensed footage, investigate explicit motion-reference support and secure rights to that performance. For a dolly movement ending on a door, a first-and-last frame pair may be more appropriate than an identity sheet.
Reference footage can contain sensitive motion data, recognizable faces and location information. Confirm rights to upload it to the selected cloud system. Avoid assuming that permission to view a dance clip gives permission to train or reproduce the dancer’s performance commercially.
Check the actual result
Build a checklist around the chosen reference’s intended job. If it was a character sheet, look for face and wardrobe drift; if a camera reference, inspect perspective and speed; if an ending frame, check the final layout. A scene that looks cinematic but does not honor the approved character or product should fail even if it would make an excellent portfolio sample.
Final integration
Store every reference with an ID, content hash, type, subject consent and model compatibility note. Don’t force a model’s unsupported input mode into a prompt by wording alone. Traditional tracking, editing and compositing remain valid alternatives when exact reference fidelity is contractually necessary.
Worked production example
A character must walk across a bridge while a crane camera follows, then finish beside a sign. If the bridge’s layout is the essential constraint, start from a scene image. If the camera path is the defining feature, use documented video motion reference on a compatible model with licensed source footage. If the character’s face is the essential constraint, a character reference mode may be necessary, but not every version offers one. Review the final shot against the specific chosen constraint rather than marking it successful merely because it looks cinematic.